On Thu, Jun 19, 2014 at 11:08:42AM +0200, Paolo Bonzini wrote:
> >static void bdrv_sync_complete(void *opaque, int ret)
> >{
> > /* do nothing. Masters do not directly interact with the backing store,
> > * only the working copy so no mutexing required.
> > */
> >}
> >
> >static void flash_sync_page(Flash *s, int page)
> >{
> > if (s->bdrv) {
> > int bdrv_sector, nb_sectors;
> > QEMUIOVector iov;
> >
> > bdrv_sector = (page * s->pi->page_size) / BDRV_SECTOR_SIZE;
> > nb_sectors = DIV_ROUND_UP(s->pi->page_size, BDRV_SECTOR_SIZE);
> > qemu_iovec_init(&iov, 1);
> > qemu_iovec_add(&iov, s->storage + bdrv_sector * BDRV_SECTOR_SIZE,
> > nb_sectors *
> > BDRV_SECTOR_SIZE);
> > bdrv_aio_writev(s->bdrv, bdrv_sector, &iov, nb_sectors,
> > bdrv_sync_complete, NULL);
> > }
> >}
>
> Using AIO is a good idea, but you could have overlapping writes here if you
> get close calls to flash_sync_page. It can be bad.
>
> Serializing can be done in fancy manners, but it can be as easy as adding
> bdrv_drain(s->bdrv) before the bdrv_aio_writev.Since this issue was present before the patch and not caused by this series I'm keeping it in the block queue. It would be nice to address this in a separate series. Stefan
pgpcYvbAl7QyW.pgp
Description: PGP signature
