On Sat, Jan 05, 2013 at 09:05:10PM +0900, Kusanagi Kouichi wrote:
> @@ -1098,15 +1101,30 @@ static int xfs_discard(BDRVRawState *s, int64_t
> sector_num, int nb_sectors)
> static coroutine_fn int raw_co_discard(BlockDriverState *bs,
> int64_t sector_num, int nb_sectors)
> {
> -#ifdef CONFIG_XFS
> +#if defined(CONFIG_FALLOCATE_PUNCH_HOLE) || defined(CONFIG_XFS)
> BDRVRawState *s = bs->opaque;
> +#endif
> + int ret = 0;
>
> +#ifdef CONFIG_FALLOCATE_PUNCH_HOLE
> + do {
> + if (fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
> + sector_num << BDRV_SECTOR_BITS,
> + nb_sectors << BDRV_SECTOR_BITS) == 0) {
> + return 0;
> + }
> + } while (errno == EINTR);
Is fallocate(fd, FALLOC_FL_PUNCH_HOLE) a blocking operation?
If yes, we need to perform this call in an aio worker thread (see
read/write/flush/ioctl in block/raw-posix.c). Failure to do this means
QEMU and the guest will be blocked during fallocate and this causes poor
performance.
Stefan
