[pbs-devel] [pve-devel] [PATCH qemu 02/11] PVE: block/pbs: fast-path reads without allocation if possible

[Wolfgang Bumiller]

On Mon, Jan 11, 2021 at 12:14:00PM +0100, Stefan Reiter wrote:
> ...and switch over to g_malloc/g_free while at it to align with other
> QEMU code.
> 
> Tracing shows the fast-path is taken almost all the time, though not
> 100% so the slow one is still necessary.

I wonder if vectored reads could be implemented directly at the
library&rust code level... this would probably have to go all the way
from here to the hyper request code though, since it's not really part
of AsyncRead :-\

> Signed-off-by: Stefan Reiter <s.reiter at proxmox.com>
> ---
> 
> For that slight unnoticable performance boost :)
> 
>  block/pbs.c | 17 ++++++++++++++---
>  1 file changed, 14 insertions(+), 3 deletions(-)
> 
> diff --git a/block/pbs.c b/block/pbs.c
> index 1481a2bfd1..fbf0d8d845 100644
> --- a/block/pbs.c
> +++ b/block/pbs.c
> @@ -200,7 +200,16 @@ static coroutine_fn int pbs_co_preadv(BlockDriverState *bs,
>      BDRVPBSState *s = bs->opaque;
>      int ret;
>      char *pbs_error = NULL;
> -    uint8_t *buf = malloc(bytes);
> +    uint8_t *buf;
> +    bool inline_buf = true;
> +
> +    /* for single-buffer IO vectors we can fast-path the write directly to it */
> +    if (qiov->niov == 1 && qiov->iov->iov_len >= bytes) {
> +        buf = qiov->iov->iov_base;
> +    } else {
> +        inline_buf = false;
> +        buf = g_malloc(bytes);
> +    }
>  
>      ReadCallbackData rcb = {
>          .co = qemu_coroutine_self(),
> @@ -218,8 +227,10 @@ static coroutine_fn int pbs_co_preadv(BlockDriverState *bs,
>          return -EIO;
>      }
>  
> -    qemu_iovec_from_buf(qiov, 0, buf, bytes);
> -    free(buf);
> +    if (!inline_buf) {
> +        qemu_iovec_from_buf(qiov, 0, buf, bytes);
> +        g_free(buf);
> +    }
>  
>      return ret;
>  }
> -- 
> 2.20.1