net/gve: fix RX buffer size alignment

  In GVE, both queue formats have RX buffer size alignment requirements
which are not respected whenever the mbuf size is greater than the
minimum required by DPDK (2048 + 128). This causes the driver to break
silently in initialization, and no queues are created, leading to no
network traffic.

This change aims to remedy this by restricting the RX receive buffer
sizes to valid sizes for their respective queue formats.

Fixes: 4bec2d0b5572 ("net/gve: support queue operations")
Fixes: 1dc00f4fc74b ("net/gve: add Rx queue setup for DQO")
Cc: junfeng.guo@intel.com
Cc: stable@dpdk.org

Signed-off-by: Joshua Washington <joshwash@google.com>
Reviewed-by: Rushil Gupta <rushilg@google.com>
---
 drivers/net/gve/gve_ethdev.c |  5 ++++-
 drivers/net/gve/gve_ethdev.h | 22 +++++++++++++++++++++-
 drivers/net/gve/gve_rx.c     | 10 +++++++++-
 drivers/net/gve/gve_rx_dqo.c |  9 ++++++++-
 4 files changed, 42 insertions(+), 4 deletions(-)
  

On 11/11/2023 12:34 AM, Joshua Washington wrote:
> In GVE, both queue formats have RX buffer size alignment requirements
> which are not respected whenever the mbuf size is greater than the
> minimum required by DPDK (2048 + 128).
>

Hi Joshua,

We don't have a way to inform application about the alignment
requirement, so drivers enforces these as you are doing in this patch.

But I am not clear with what is "minimum required by DPDK", since
application can provide smaller mbufs.
Also not clear why this alignment cause problem only with mbuf size
bigger than 2048 + 128 bytes. Can you please clarify?

> This causes the driver to break
> silently in initialization, and no queues are created, leading to no
> network traffic.
> 
> This change aims to remedy this by restricting the RX receive buffer
> sizes to valid sizes for their respective queue formats.
> 
> Fixes: 4bec2d0b5572 ("net/gve: support queue operations")
> Fixes: 1dc00f4fc74b ("net/gve: add Rx queue setup for DQO")
> Cc: junfeng.guo@intel.com
> Cc: stable@dpdk.org
> 
> Signed-off-by: Joshua Washington <joshwash@google.com>
> Reviewed-by: Rushil Gupta <rushilg@google.com>

<...>

> @@ -337,6 +343,20 @@ gve_clear_device_rings_ok(struct gve_priv *priv)
>  				&priv->state_flags);
>  }
>  
> +static inline int
> +gve_validate_rx_buffer_size(struct gve_priv *priv, uint16_t rx_buffer_size)
> +{
> +	uint16_t min_rx_buffer_size = gve_is_gqi(priv) ?
> +		GVE_RX_MIN_BUF_SIZE_GQI : GVE_RX_MIN_BUF_SIZE_DQO;
> +	if (rx_buffer_size < min_rx_buffer_size) {
> +		PMD_DRV_LOG(ERR, "mbuf size must be at least %hu bytes",
> +			    min_rx_buffer_size);
> +		return -EINVAL;
> +	}
> +
>

When 'dev_info->min_rx_bufsize' set correctly, above check should be
done in ethdev level, can you please check 'rte_eth_check_rx_mempool()'.


> +	return 0;
> +}
> +
>  int
>  gve_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_id, uint16_t nb_desc,
>  		   unsigned int socket_id, const struct rte_eth_rxconf *conf,
> diff --git a/drivers/net/gve/gve_rx.c b/drivers/net/gve/gve_rx.c
> index b8c92ccda0..0049c6428d 100644
> --- a/drivers/net/gve/gve_rx.c
> +++ b/drivers/net/gve/gve_rx.c
> @@ -301,6 +301,7 @@ gve_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_id,
>  	const struct rte_memzone *mz;
>  	struct gve_rx_queue *rxq;
>  	uint16_t free_thresh;
> +	uint32_t mbuf_len;
>  	int err = 0;
>  
>  	if (nb_desc != hw->rx_desc_cnt) {
> @@ -344,7 +345,14 @@ gve_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_id,
>  	rxq->hw = hw;
>  	rxq->ntfy_addr = &hw->db_bar2[rte_be_to_cpu_32(hw->irq_dbs[rxq->ntfy_id].id)];
>  
> -	rxq->rx_buf_len = rte_pktmbuf_data_room_size(rxq->mpool) - RTE_PKTMBUF_HEADROOM;
> +	mbuf_len =
> +		rte_pktmbuf_data_room_size(rxq->mpool) - RTE_PKTMBUF_HEADROOM;
> +	err = gve_validate_rx_buffer_size(hw, mbuf_len);
> +	if (err)
> +		goto err_rxq;
> +	rxq->rx_buf_len =
> +		RTE_MIN((uint16_t)GVE_RX_MAX_BUF_SIZE_GQI,
> +			RTE_ALIGN_FLOOR(mbuf_len, GVE_RX_BUF_ALIGN_GQI));
>

Just for your info, this release 'dev_info.max_rx_bufsize' and ethdev
layer note added [1] if user provides mbuf size bigger than this value.
Ethdev layer not is mainly for memmory optimization, but above check is
required for driver.

[1]
https://git.dpdk.org/dpdk/commit/?id=75c7849a9dcca356985fdb87f2d11cae135dfb1a
  

Hello Ferruh,

But I am not clear with what is "minimum required by DPDK", since
> application can provide smaller mbufs.
> Also not clear why this alignment cause problem only with mbuf size
> bigger than 2048 + 128 bytes. Can you please clarify?
>

My apologies, the statement "minimum required by DPDK" is a typo; it should
probably say "minimum recommended", per
https://doc.dpdk.org/api/rte__mbuf__core_8h.html#a185c46bcdbfa90f6c50a4b037a93313f.
The GVE GQ driver is one which requires a packet buffer size of at least
2K. This alignment issue manifests in a different way when the mbuf size is
smaller than the minimum supported by the device, but it is an issue
nonetheless. I will fix the wording in the commit description in an updated
patch.

When 'dev_info->min_rx_bufsize' set correctly, above check should be
> done in ethdev level, can you please check 'rte_eth_check_rx_mempool()'.
>

This validation path does seem to be hit when running testpmd:

# dpdk-testpmd -- -a --stats-period=1 --forward-mode=txonly --rxq=$N
--txq=$N --nb-cores=$(($N + 1)) --mbuf-size=10
...
mb_pool_0 mbuf_data_room_size 10 < 1152 (128 + 1024)
Fail to configure port 0 rx queues
Port 0 is closed
EAL: Error - exiting with code: 1
  Cause: Start ports failed

I can remove this check from the driver, as it is redundant.

Just for your info, this release 'dev_info.max_rx_bufsize' and ethdev
> layer note added [1] if user provides mbuf size bigger than this value.
> Ethdev layer not is mainly for memmory optimization, but above check is
> required for driver.
>
> [1]
>
> https://git.dpdk.org/dpdk/commit/?id=75c7849a9dcca356985fdb87f2d11cae135dfb1a


If I were to add GVE support for max buffer size to this patch, how would
that interact with backports? Is it possible to include only parts of a
patch in a backport?



On Fri, Nov 10, 2023 at 8:18 PM Ferruh Yigit <ferruh.yigit@amd.com> wrote:

> On 11/11/2023 12:34 AM, Joshua Washington wrote:
> > In GVE, both queue formats have RX buffer size alignment requirements
> > which are not respected whenever the mbuf size is greater than the
> > minimum required by DPDK (2048 + 128).
> >
>
> Hi Joshua,
>
> We don't have a way to inform application about the alignment
> requirement, so drivers enforces these as you are doing in this patch.
>
> But I am not clear with what is "minimum required by DPDK", since
> application can provide smaller mbufs.
> Also not clear why this alignment cause problem only with mbuf size
> bigger than 2048 + 128 bytes. Can you please clarify?
>
> > This causes the driver to break
> > silently in initialization, and no queues are created, leading to no
> > network traffic.
> >
> > This change aims to remedy this by restricting the RX receive buffer
> > sizes to valid sizes for their respective queue formats.
> >
> > Fixes: 4bec2d0b5572 ("net/gve: support queue operations")
> > Fixes: 1dc00f4fc74b ("net/gve: add Rx queue setup for DQO")
> > Cc: junfeng.guo@intel.com
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Joshua Washington <joshwash@google.com>
> > Reviewed-by: Rushil Gupta <rushilg@google.com>
>
> <...>
>
> > @@ -337,6 +343,20 @@ gve_clear_device_rings_ok(struct gve_priv *priv)
> >                               &priv->state_flags);
> >  }
> >
> > +static inline int
> > +gve_validate_rx_buffer_size(struct gve_priv *priv, uint16_t
> rx_buffer_size)
> > +{
> > +     uint16_t min_rx_buffer_size = gve_is_gqi(priv) ?
> > +             GVE_RX_MIN_BUF_SIZE_GQI : GVE_RX_MIN_BUF_SIZE_DQO;
> > +     if (rx_buffer_size < min_rx_buffer_size) {
> > +             PMD_DRV_LOG(ERR, "mbuf size must be at least %hu bytes",
> > +                         min_rx_buffer_size);
> > +             return -EINVAL;
> > +     }
> > +
> >
>
> When 'dev_info->min_rx_bufsize' set correctly, above check should be
> done in ethdev level, can you please check 'rte_eth_check_rx_mempool()'.
>
>
> > +     return 0;
> > +}
> > +
> >  int
> >  gve_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_id, uint16_t
> nb_desc,
> >                  unsigned int socket_id, const struct rte_eth_rxconf
> *conf,
> > diff --git a/drivers/net/gve/gve_rx.c b/drivers/net/gve/gve_rx.c
> > index b8c92ccda0..0049c6428d 100644
> > --- a/drivers/net/gve/gve_rx.c
> > +++ b/drivers/net/gve/gve_rx.c
> > @@ -301,6 +301,7 @@ gve_rx_queue_setup(struct rte_eth_dev *dev, uint16_t
> queue_id,
> >       const struct rte_memzone *mz;
> >       struct gve_rx_queue *rxq;
> >       uint16_t free_thresh;
> > +     uint32_t mbuf_len;
> >       int err = 0;
> >
> >       if (nb_desc != hw->rx_desc_cnt) {
> > @@ -344,7 +345,14 @@ gve_rx_queue_setup(struct rte_eth_dev *dev,
> uint16_t queue_id,
> >       rxq->hw = hw;
> >       rxq->ntfy_addr =
> &hw->db_bar2[rte_be_to_cpu_32(hw->irq_dbs[rxq->ntfy_id].id)];
> >
> > -     rxq->rx_buf_len = rte_pktmbuf_data_room_size(rxq->mpool) -
> RTE_PKTMBUF_HEADROOM;
> > +     mbuf_len =
> > +             rte_pktmbuf_data_room_size(rxq->mpool) -
> RTE_PKTMBUF_HEADROOM;
> > +     err = gve_validate_rx_buffer_size(hw, mbuf_len);
> > +     if (err)
> > +             goto err_rxq;
> > +     rxq->rx_buf_len =
> > +             RTE_MIN((uint16_t)GVE_RX_MAX_BUF_SIZE_GQI,
> > +                     RTE_ALIGN_FLOOR(mbuf_len, GVE_RX_BUF_ALIGN_GQI));
> >
>
> Just for your info, this release 'dev_info.max_rx_bufsize' and ethdev
> layer note added [1] if user provides mbuf size bigger than this value.
> Ethdev layer not is mainly for memmory optimization, but above check is
> required for driver.
>
> [1]
>
> https://git.dpdk.org/dpdk/commit/?id=75c7849a9dcca356985fdb87f2d11cae135dfb1a
>
>