Virtualization - Feb 2021 - [PATCH RFC v2 03/10] vringh: reset kiov 'consumed' field in __vringh_iov()

On 2021/1/28 ??10:41, Stefano Garzarella wrote:
> __vringh_iov() overwrites the contents of riov and wiov, in fact it
> resets the 'i' and 'used' fields, but also the consumed
field should
> be reset to avoid an inconsistent state.
>
> Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>

I had a question(I remember we had some discussion like this but I 
forget the conclusion):

I see e.g in vringh_getdesc_kern() it has the following comment:

/*
 ?* Note that you may need to clean up riov and wiov, even on error!
 ?*/

So it looks to me the correct way is to call vringh_kiov_cleanup() before?

Thanks

> ---
>   drivers/vhost/vringh.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/vhost/vringh.c b/drivers/vhost/vringh.c
> index f68122705719..bee63d68201a 100644
> --- a/drivers/vhost/vringh.c
> +++ b/drivers/vhost/vringh.c
> @@ -290,9 +290,9 @@ __vringh_iov(struct vringh *vrh, u16 i,
>   		return -EINVAL;
>   
>   	if (riov)
> -		riov->i = riov->used = 0;
> +		riov->i = riov->used = riov->consumed = 0;
>   	if (wiov)
> -		wiov->i = wiov->used = 0;
> +		wiov->i = wiov->used = wiov->consumed = 0;
>   
>   	for (;;) {
>   		void *addr;

On Mon, Feb 01, 2021 at 01:40:01PM +0800, Jason Wang
wrote:
>
>On 2021/1/28 ??10:41, Stefano Garzarella wrote:
>>__vringh_iov() overwrites the contents of riov and wiov, in fact it
>>resets the 'i' and 'used' fields, but also the consumed
field should
>>be reset to avoid an inconsistent state.
>>
>>Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
>
>
>I had a question(I remember we had some discussion like this but I 
>forget the conclusion):
Sorry, I forgot to update you.
>
>I see e.g in vringh_getdesc_kern() it has the following comment:
>
>/*
>?* Note that you may need to clean up riov and wiov, even on error!
>?*/
>
>So it looks to me the correct way is to call vringh_kiov_cleanup() 
>before?
Looking at the code the right pattern should be:

     vringh_getdesc_*(..., &out_iov, &in_iov, ...);

     // use out_iov and in_iov

     vringh_kiov_cleanup(&out_iov);
     vringh_kiov_cleanup(&in_iov);

This because vringh_getdesc_*() calls __vringh_iov() where 
resize_iovec() is called to allocate the iov wrapped by 'struct 
vringh_kiov' and vringh_kiov_cleanup() frees that memory.

Looking better, __vringh_iov() is able to extend a 'vringh_kiov' 
pre-allocated, so in order to avoid to allocate and free the iov for 
each request we can avoid to call vringh_kiov_cleanup(), but this patch 
is needed to avoid an inconsistent state.

And also patch "vdpa_sim: cleanup kiovs in vdpasim_free()" is required
to free the iov when the device is going away.

Does that make sense to you?

Maybe I should add a comment in vringh.c to explain this better.

Thanks,
Stefano
>
>Thanks
>
>
>>---
>>  drivers/vhost/vringh.c | 4 ++--
>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>
>>diff --git a/drivers/vhost/vringh.c b/drivers/vhost/vringh.c
>>index f68122705719..bee63d68201a 100644
>>--- a/drivers/vhost/vringh.c
>>+++ b/drivers/vhost/vringh.c
>>@@ -290,9 +290,9 @@ __vringh_iov(struct vringh *vrh, u16 i,
>>  		return -EINVAL;
>>  	if (riov)
>>-		riov->i = riov->used = 0;
>>+		riov->i = riov->used = riov->consumed = 0;
>>  	if (wiov)
>>-		wiov->i = wiov->used = 0;
>>+		wiov->i = wiov->used = wiov->consumed = 0;
>>  	for (;;) {
>>  		void *addr;
>