[dpdk-dev,3/3] net/mlx5: adjust removal error
Checks
Commit Message
Fail-safe PMD expects to get -ENODEV error value if sub PMD control
command fails because of device removal.
Make control callbacks return with -ENODEV when the device has
disappeared.
Signed-off-by: Matan Azrad <matan@mellanox.com>
---
drivers/net/mlx5/mlx5.h | 1 +
drivers/net/mlx5/mlx5_ethdev.c | 39 +++++++++++++++++++++++++++++++++++----
drivers/net/mlx5/mlx5_flow.c | 2 ++
drivers/net/mlx5/mlx5_rss.c | 4 ++++
drivers/net/mlx5/mlx5_rxq.c | 12 ++++++++++--
drivers/net/mlx5/mlx5_stats.c | 6 +++++-
drivers/net/mlx5/mlx5_txq.c | 2 ++
7 files changed, 59 insertions(+), 7 deletions(-)
Comments
On Thu, Nov 02, 2017 at 03:42:04PM +0000, Matan Azrad wrote:
> Fail-safe PMD expects to get -ENODEV error value if sub PMD control
> command fails because of device removal.
>
> Make control callbacks return with -ENODEV when the device has
> disappeared.
>
> Signed-off-by: Matan Azrad <matan@mellanox.com>
In short I have the same comments as on the mlx4 patch about usage
consistency, this also applies to mlx5; mlx5_removed() should be only used
by the public callbacks from struct eth_dev_ops.
There's an additional difficulty with this PMD, you need to take into
account the fact it provides secondary process support (mlx5_dev_sec_ops).
I think secondary processes do not have any IBV context available for
mlx5_removed() to query, which should resolve to a no-op in this case.
Make sure secondary processes do not crash whatever happens.
See below for coding style and other issues.
> ---
> drivers/net/mlx5/mlx5.h | 1 +
> drivers/net/mlx5/mlx5_ethdev.c | 39 +++++++++++++++++++++++++++++++++++----
> drivers/net/mlx5/mlx5_flow.c | 2 ++
> drivers/net/mlx5/mlx5_rss.c | 4 ++++
> drivers/net/mlx5/mlx5_rxq.c | 12 ++++++++++--
> drivers/net/mlx5/mlx5_stats.c | 6 +++++-
> drivers/net/mlx5/mlx5_txq.c | 2 ++
> 7 files changed, 59 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
> index e6a69b8..0dd104a 100644
> --- a/drivers/net/mlx5/mlx5.h
> +++ b/drivers/net/mlx5/mlx5.h
> @@ -208,6 +208,7 @@ int mlx5_ibv_device_to_pci_addr(const struct ibv_device *,
> int mlx5_set_link_up(struct rte_eth_dev *dev);
> void priv_dev_select_tx_function(struct priv *priv, struct rte_eth_dev *dev);
> void priv_dev_select_rx_function(struct priv *priv, struct rte_eth_dev *dev);
> +int mlx5_removed(const struct priv *priv);
>
> /* mlx5_mac.c */
>
> diff --git a/drivers/net/mlx5/mlx5_ethdev.c b/drivers/net/mlx5/mlx5_ethdev.c
> index c31ea4b..bf61cd6 100644
> --- a/drivers/net/mlx5/mlx5_ethdev.c
> +++ b/drivers/net/mlx5/mlx5_ethdev.c
> @@ -394,6 +394,8 @@ struct priv *
>
> ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
> if (ret == -1) {
> + if (mlx5_removed(priv))
> + errno = ENODEV;
> DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
> name, value_str, value, strerror(errno));
> return -1;
> @@ -925,13 +927,17 @@ struct priv *
> {
> struct utsname utsname;
> int ver[3];
> + int ret;
>
> if (uname(&utsname) == -1 ||
> sscanf(utsname.release, "%d.%d.%d",
> &ver[0], &ver[1], &ver[2]) != 3 ||
> KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
> - return mlx5_link_update_unlocked_gset(dev, wait_to_complete);
> - return mlx5_link_update_unlocked_gs(dev, wait_to_complete);
> + ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
> + ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
Besides the extra space after "ret =", I think this doesn't work as
intended. A "else" statement is necessary.
> + if (ret && mlx5_removed(mlx5_get_priv(dev)))
> + return -ENODEV;
> + return ret;
> }
>
> /**
> @@ -978,6 +984,8 @@ struct priv *
> strerror(ret));
> priv_unlock(priv);
> assert(ret >= 0);
> + if (mlx5_removed(priv))
> + return -ENODEV;
> return -ret;
> }
>
> @@ -1029,6 +1037,8 @@ struct priv *
> out:
> priv_unlock(priv);
> assert(ret >= 0);
> + if (mlx5_removed(priv))
> + return -ENODEV;
> return -ret;
> }
>
> @@ -1083,6 +1093,8 @@ struct priv *
> out:
> priv_unlock(priv);
> assert(ret >= 0);
> + if (mlx5_removed(priv))
> + return -ENODEV;
> return -ret;
> }
>
> @@ -1364,13 +1376,13 @@ struct priv *
> if (up) {
> err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
> if (err)
> - return err;
> + return errno == ENODEV ? -ENODEV : err;
There is a documentation issue here since the mlx5 PMD didn't get all the
errno consistency fixes that mlx4 got, however err is documented as being -1
in case of error, whereas priv_dev_set_link() returns a positive errno value
instead and mlx5_set_link_down/up() should return only negative errno values
but are documented as returning positive ones.
Anyway to keep it short: currently in mlx5, priv_*() => positive errno and
the public-facing mlx5_*() => negative errno, hence you should return a
positive ENODEV here.
You could avoid this mess by patching the public callbacks only and not
internal functions like this one.
> priv_dev_select_tx_function(priv, dev);
> priv_dev_select_rx_function(priv, dev);
> } else {
> err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
> if (err)
> - return err;
> + return errno == ENODEV ? -ENODEV : err;
Same here.
> dev->rx_pkt_burst = removed_rx_burst;
> dev->tx_pkt_burst = removed_tx_burst;
> }
> @@ -1474,3 +1486,22 @@ struct priv *
> dev->rx_pkt_burst = mlx5_rx_burst;
> }
> }
> +
> +/**
> + * Check if mlx5 device was removed.
> + *
"mlx5" is redundant.
As with mlx4, a short paragraph should describe where this function is
supposed to be used.
> + * @param priv
> + * Pointer to private structure.
> + *
> + * @return
> + * -ENODEV when device is removed and rte_errno is set, otherwise 0.
> + */
> +int
> +mlx5_removed(const struct priv *priv)
> +{
> + struct ibv_device_attr device_attr;
> +
> + if (ibv_query_device(priv->ctx, &device_attr) == EIO)
> + return -(rte_errno = ENODEV);
Coding rules prohibit this kind of affectation, see mlx4 comments.
> + return 0;
> +}
> diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
> index 5f49bf5..448c0a3 100644
> --- a/drivers/net/mlx5/mlx5_flow.c
> +++ b/drivers/net/mlx5/mlx5_flow.c
> @@ -3068,6 +3068,8 @@ struct rte_flow *
> priv_lock(priv);
> ret = priv_fdir_ctrl_func(priv, filter_op, arg);
> priv_unlock(priv);
> + if (ret && mlx5_removed(priv))
> + ret = ENODEV;
> break;
> default:
> ERROR("%p: filter type (%d) not supported",
> diff --git a/drivers/net/mlx5/mlx5_rss.c b/drivers/net/mlx5/mlx5_rss.c
> index f3de46d..1ad9269 100644
> --- a/drivers/net/mlx5/mlx5_rss.c
> +++ b/drivers/net/mlx5/mlx5_rss.c
> @@ -250,6 +250,8 @@
> priv_lock(priv);
> ret = priv_dev_rss_reta_query(priv, reta_conf, reta_size);
> priv_unlock(priv);
> + if (ret && mlx5_removed(priv))
> + return -ENODEV;
> return -ret;
> }
>
> @@ -282,5 +284,7 @@
> mlx5_dev_stop(dev);
> mlx5_dev_start(dev);
> }
> + if (ret && mlx5_removed(priv))
> + return -ENODEV;
> return -ret;
> }
> diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
> index a1f382b..c9a549d 100644
> --- a/drivers/net/mlx5/mlx5_rxq.c
> +++ b/drivers/net/mlx5/mlx5_rxq.c
> @@ -278,6 +278,8 @@
> (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
> out:
> priv_unlock(priv);
> + if (mlx5_removed(priv))
> + return -ENODEV;
> return -ret;
> }
>
> @@ -485,8 +487,11 @@
> }
> exit:
> priv_unlock(priv);
> - if (ret)
> + if (ret) {
> WARN("unable to arm interrupt on rx queue %d", rx_queue_id);
> + if (mlx5_removed(priv))
> + return -ENODEV;
> + }
> return -ret;
> }
>
> @@ -537,9 +542,12 @@
> if (rxq_ibv)
> mlx5_priv_rxq_ibv_release(priv, rxq_ibv);
> priv_unlock(priv);
> - if (ret)
> + if (ret) {
> WARN("unable to disable interrupt on rx queue %d",
> rx_queue_id);
> + if (mlx5_removed(priv))
> + return -ENODEV;
> + }
> return -ret;
> }
>
> diff --git a/drivers/net/mlx5/mlx5_stats.c b/drivers/net/mlx5/mlx5_stats.c
> index 5e225d3..33b2a60 100644
> --- a/drivers/net/mlx5/mlx5_stats.c
> +++ b/drivers/net/mlx5/mlx5_stats.c
> @@ -438,13 +438,17 @@ struct mlx5_counter_ctrl {
> stats_n = priv_ethtool_get_stats_n(priv);
> if (stats_n < 0) {
> priv_unlock(priv);
> - return -1;
> + ret = -1;
> + goto error;
> }
> if (xstats_ctrl->stats_n != stats_n)
> priv_xstats_init(priv);
> ret = priv_xstats_get(priv, stats);
> priv_unlock(priv);
> }
> +error:
> + if (ret < 0 && mlx5_removed(priv))
> + return -ENODEV;
> return ret;
> }
>
> diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c
> index fbb2630..a0101cb 100644
> --- a/drivers/net/mlx5/mlx5_txq.c
> +++ b/drivers/net/mlx5/mlx5_txq.c
> @@ -186,6 +186,8 @@
> (*priv->txqs)[idx] = &txq_ctrl->txq;
> out:
> priv_unlock(priv);
> + if (mlx5_removed(priv))
> + return -ENODEV;
> return -ret;
> }
>
> --
> 1.8.3.1
>
Hi Adrien,
Thanks for this too.
> -----Original Message-----
> From: Adrien Mazarguil [mailto:adrien.mazarguil@6wind.com]
> Sent: Friday, November 3, 2017 3:06 PM
> To: Matan Azrad <matan@mellanox.com>
> Cc: Gaetan Rivet <gaetan.rivet@6wind.com>; dev@dpdk.org
> Subject: Re: [PATCH 3/3] net/mlx5: adjust removal error
>
> On Thu, Nov 02, 2017 at 03:42:04PM +0000, Matan Azrad wrote:
> > Fail-safe PMD expects to get -ENODEV error value if sub PMD control
> > command fails because of device removal.
> >
> > Make control callbacks return with -ENODEV when the device has
> > disappeared.
> >
> > Signed-off-by: Matan Azrad <matan@mellanox.com>
>
> In short I have the same comments as on the mlx4 patch about usage
> consistency, this also applies to mlx5; mlx5_removed() should be only used
> by the public callbacks from struct eth_dev_ops.
>
> There's an additional difficulty with this PMD, you need to take into account
> the fact it provides secondary process support (mlx5_dev_sec_ops).
> I think secondary processes do not have any IBV context available for
> mlx5_removed() to query, which should resolve to a no-op in this case.
> Make sure secondary processes do not crash whatever happens.
>
Will check it, thanks!
> See below for coding style and other issues.
>
> > ---
> > drivers/net/mlx5/mlx5.h | 1 +
> > drivers/net/mlx5/mlx5_ethdev.c | 39
> +++++++++++++++++++++++++++++++++++----
> > drivers/net/mlx5/mlx5_flow.c | 2 ++
> > drivers/net/mlx5/mlx5_rss.c | 4 ++++
> > drivers/net/mlx5/mlx5_rxq.c | 12 ++++++++++--
> > drivers/net/mlx5/mlx5_stats.c | 6 +++++-
> > drivers/net/mlx5/mlx5_txq.c | 2 ++
> > 7 files changed, 59 insertions(+), 7 deletions(-)
> >
> > diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h index
> > e6a69b8..0dd104a 100644
> > --- a/drivers/net/mlx5/mlx5.h
> > +++ b/drivers/net/mlx5/mlx5.h
> > @@ -208,6 +208,7 @@ int mlx5_ibv_device_to_pci_addr(const struct
> > ibv_device *, int mlx5_set_link_up(struct rte_eth_dev *dev); void
> > priv_dev_select_tx_function(struct priv *priv, struct rte_eth_dev
> > *dev); void priv_dev_select_rx_function(struct priv *priv, struct
> > rte_eth_dev *dev);
> > +int mlx5_removed(const struct priv *priv);
> >
> > /* mlx5_mac.c */
> >
> > diff --git a/drivers/net/mlx5/mlx5_ethdev.c
> > b/drivers/net/mlx5/mlx5_ethdev.c index c31ea4b..bf61cd6 100644
> > --- a/drivers/net/mlx5/mlx5_ethdev.c
> > +++ b/drivers/net/mlx5/mlx5_ethdev.c
> > @@ -394,6 +394,8 @@ struct priv *
> >
> > ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
> > if (ret == -1) {
> > + if (mlx5_removed(priv))
> > + errno = ENODEV;
> > DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
> > name, value_str, value, strerror(errno));
> > return -1;
> > @@ -925,13 +927,17 @@ struct priv *
> > {
> > struct utsname utsname;
> > int ver[3];
> > + int ret;
> >
> > if (uname(&utsname) == -1 ||
> > sscanf(utsname.release, "%d.%d.%d",
> > &ver[0], &ver[1], &ver[2]) != 3 ||
> > KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9,
> 0))
> > - return mlx5_link_update_unlocked_gset(dev,
> wait_to_complete);
> > - return mlx5_link_update_unlocked_gs(dev, wait_to_complete);
> > + ret = mlx5_link_update_unlocked_gset(dev,
> wait_to_complete);
> > + ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
>
> Besides the extra space after "ret =", I think this doesn't work as intended. A
> "else" statement is necessary.
>
Will fix it, thanks!
> > + if (ret && mlx5_removed(mlx5_get_priv(dev)))
> > + return -ENODEV;
> > + return ret;
> > }
> >
> > /**
> > @@ -978,6 +984,8 @@ struct priv *
> > strerror(ret));
> > priv_unlock(priv);
> > assert(ret >= 0);
> > + if (mlx5_removed(priv))
> > + return -ENODEV;
> > return -ret;
> > }
> >
> > @@ -1029,6 +1037,8 @@ struct priv *
> > out:
> > priv_unlock(priv);
> > assert(ret >= 0);
> > + if (mlx5_removed(priv))
> > + return -ENODEV;
> > return -ret;
> > }
> >
> > @@ -1083,6 +1093,8 @@ struct priv *
> > out:
> > priv_unlock(priv);
> > assert(ret >= 0);
> > + if (mlx5_removed(priv))
> > + return -ENODEV;
> > return -ret;
> > }
> >
> > @@ -1364,13 +1376,13 @@ struct priv *
> > if (up) {
> > err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
> > if (err)
> > - return err;
> > + return errno == ENODEV ? -ENODEV : err;
>
> There is a documentation issue here since the mlx5 PMD didn't get all the
> errno consistency fixes that mlx4 got, however err is documented as being -1
> in case of error, whereas priv_dev_set_link() returns a positive errno value
> instead and mlx5_set_link_down/up() should return only negative errno
> values but are documented as returning positive ones.
>
> Anyway to keep it short: currently in mlx5, priv_*() => positive errno and the
> public-facing mlx5_*() => negative errno, hence you should return a positive
> ENODEV here.
>
Ok.
> You could avoid this mess by patching the public callbacks only and not
> internal functions like this one.
>
> > priv_dev_select_tx_function(priv, dev);
> > priv_dev_select_rx_function(priv, dev);
> > } else {
> > err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
> > if (err)
> > - return err;
> > + return errno == ENODEV ? -ENODEV : err;
>
> Same here.
>
> > dev->rx_pkt_burst = removed_rx_burst;
> > dev->tx_pkt_burst = removed_tx_burst;
> > }
> > @@ -1474,3 +1486,22 @@ struct priv *
> > dev->rx_pkt_burst = mlx5_rx_burst;
> > }
> > }
> > +
> > +/**
> > + * Check if mlx5 device was removed.
> > + *
>
> "mlx5" is redundant.
>
> As with mlx4, a short paragraph should describe where this function is
> supposed to be used.
>
> > + * @param priv
> > + * Pointer to private structure.
> > + *
> > + * @return
> > + * -ENODEV when device is removed and rte_errno is set, otherwise 0.
> > + */
> > +int
> > +mlx5_removed(const struct priv *priv) {
> > + struct ibv_device_attr device_attr;
> > +
> > + if (ibv_query_device(priv->ctx, &device_attr) == EIO)
> > + return -(rte_errno = ENODEV);
>
> Coding rules prohibit this kind of affectation, see mlx4 comments.
>
> > + return 0;
> > +}
> > diff --git a/drivers/net/mlx5/mlx5_flow.c
> > b/drivers/net/mlx5/mlx5_flow.c index 5f49bf5..448c0a3 100644
> > --- a/drivers/net/mlx5/mlx5_flow.c
> > +++ b/drivers/net/mlx5/mlx5_flow.c
> > @@ -3068,6 +3068,8 @@ struct rte_flow *
> > priv_lock(priv);
> > ret = priv_fdir_ctrl_func(priv, filter_op, arg);
> > priv_unlock(priv);
> > + if (ret && mlx5_removed(priv))
> > + ret = ENODEV;
> > break;
> > default:
> > ERROR("%p: filter type (%d) not supported", diff --git
> > a/drivers/net/mlx5/mlx5_rss.c b/drivers/net/mlx5/mlx5_rss.c index
> > f3de46d..1ad9269 100644
> > --- a/drivers/net/mlx5/mlx5_rss.c
> > +++ b/drivers/net/mlx5/mlx5_rss.c
> > @@ -250,6 +250,8 @@
> > priv_lock(priv);
> > ret = priv_dev_rss_reta_query(priv, reta_conf, reta_size);
> > priv_unlock(priv);
> > + if (ret && mlx5_removed(priv))
> > + return -ENODEV;
> > return -ret;
> > }
> >
> > @@ -282,5 +284,7 @@
> > mlx5_dev_stop(dev);
> > mlx5_dev_start(dev);
> > }
> > + if (ret && mlx5_removed(priv))
> > + return -ENODEV;
> > return -ret;
> > }
> > diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
> > index a1f382b..c9a549d 100644
> > --- a/drivers/net/mlx5/mlx5_rxq.c
> > +++ b/drivers/net/mlx5/mlx5_rxq.c
> > @@ -278,6 +278,8 @@
> > (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
> > out:
> > priv_unlock(priv);
> > + if (mlx5_removed(priv))
> > + return -ENODEV;
> > return -ret;
> > }
> >
> > @@ -485,8 +487,11 @@
> > }
> > exit:
> > priv_unlock(priv);
> > - if (ret)
> > + if (ret) {
> > WARN("unable to arm interrupt on rx queue %d",
> rx_queue_id);
> > + if (mlx5_removed(priv))
> > + return -ENODEV;
> > + }
> > return -ret;
> > }
> >
> > @@ -537,9 +542,12 @@
> > if (rxq_ibv)
> > mlx5_priv_rxq_ibv_release(priv, rxq_ibv);
> > priv_unlock(priv);
> > - if (ret)
> > + if (ret) {
> > WARN("unable to disable interrupt on rx queue %d",
> > rx_queue_id);
> > + if (mlx5_removed(priv))
> > + return -ENODEV;
> > + }
> > return -ret;
> > }
> >
> > diff --git a/drivers/net/mlx5/mlx5_stats.c
> > b/drivers/net/mlx5/mlx5_stats.c index 5e225d3..33b2a60 100644
> > --- a/drivers/net/mlx5/mlx5_stats.c
> > +++ b/drivers/net/mlx5/mlx5_stats.c
> > @@ -438,13 +438,17 @@ struct mlx5_counter_ctrl {
> > stats_n = priv_ethtool_get_stats_n(priv);
> > if (stats_n < 0) {
> > priv_unlock(priv);
> > - return -1;
> > + ret = -1;
> > + goto error;
> > }
> > if (xstats_ctrl->stats_n != stats_n)
> > priv_xstats_init(priv);
> > ret = priv_xstats_get(priv, stats);
> > priv_unlock(priv);
> > }
> > +error:
> > + if (ret < 0 && mlx5_removed(priv))
> > + return -ENODEV;
> > return ret;
> > }
> >
> > diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c
> > index fbb2630..a0101cb 100644
> > --- a/drivers/net/mlx5/mlx5_txq.c
> > +++ b/drivers/net/mlx5/mlx5_txq.c
> > @@ -186,6 +186,8 @@
> > (*priv->txqs)[idx] = &txq_ctrl->txq;
> > out:
> > priv_unlock(priv);
> > + if (mlx5_removed(priv))
> > + return -ENODEV;
> > return -ret;
> > }
> >
> > --
> > 1.8.3.1
> >
>
> --
> Adrien Mazarguil
> 6WIND
@@ -208,6 +208,7 @@ int mlx5_ibv_device_to_pci_addr(const struct ibv_device *,
int mlx5_set_link_up(struct rte_eth_dev *dev);
void priv_dev_select_tx_function(struct priv *priv, struct rte_eth_dev *dev);
void priv_dev_select_rx_function(struct priv *priv, struct rte_eth_dev *dev);
+int mlx5_removed(const struct priv *priv);
/* mlx5_mac.c */
@@ -394,6 +394,8 @@ struct priv *
ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
if (ret == -1) {
+ if (mlx5_removed(priv))
+ errno = ENODEV;
DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
name, value_str, value, strerror(errno));
return -1;
@@ -925,13 +927,17 @@ struct priv *
{
struct utsname utsname;
int ver[3];
+ int ret;
if (uname(&utsname) == -1 ||
sscanf(utsname.release, "%d.%d.%d",
&ver[0], &ver[1], &ver[2]) != 3 ||
KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
- return mlx5_link_update_unlocked_gset(dev, wait_to_complete);
- return mlx5_link_update_unlocked_gs(dev, wait_to_complete);
+ ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
+ ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
+ if (ret && mlx5_removed(mlx5_get_priv(dev)))
+ return -ENODEV;
+ return ret;
}
/**
@@ -978,6 +984,8 @@ struct priv *
strerror(ret));
priv_unlock(priv);
assert(ret >= 0);
+ if (mlx5_removed(priv))
+ return -ENODEV;
return -ret;
}
@@ -1029,6 +1037,8 @@ struct priv *
out:
priv_unlock(priv);
assert(ret >= 0);
+ if (mlx5_removed(priv))
+ return -ENODEV;
return -ret;
}
@@ -1083,6 +1093,8 @@ struct priv *
out:
priv_unlock(priv);
assert(ret >= 0);
+ if (mlx5_removed(priv))
+ return -ENODEV;
return -ret;
}
@@ -1364,13 +1376,13 @@ struct priv *
if (up) {
err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
if (err)
- return err;
+ return errno == ENODEV ? -ENODEV : err;
priv_dev_select_tx_function(priv, dev);
priv_dev_select_rx_function(priv, dev);
} else {
err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
if (err)
- return err;
+ return errno == ENODEV ? -ENODEV : err;
dev->rx_pkt_burst = removed_rx_burst;
dev->tx_pkt_burst = removed_tx_burst;
}
@@ -1474,3 +1486,22 @@ struct priv *
dev->rx_pkt_burst = mlx5_rx_burst;
}
}
+
+/**
+ * Check if mlx5 device was removed.
+ *
+ * @param priv
+ * Pointer to private structure.
+ *
+ * @return
+ * -ENODEV when device is removed and rte_errno is set, otherwise 0.
+ */
+int
+mlx5_removed(const struct priv *priv)
+{
+ struct ibv_device_attr device_attr;
+
+ if (ibv_query_device(priv->ctx, &device_attr) == EIO)
+ return -(rte_errno = ENODEV);
+ return 0;
+}
@@ -3068,6 +3068,8 @@ struct rte_flow *
priv_lock(priv);
ret = priv_fdir_ctrl_func(priv, filter_op, arg);
priv_unlock(priv);
+ if (ret && mlx5_removed(priv))
+ ret = ENODEV;
break;
default:
ERROR("%p: filter type (%d) not supported",
@@ -250,6 +250,8 @@
priv_lock(priv);
ret = priv_dev_rss_reta_query(priv, reta_conf, reta_size);
priv_unlock(priv);
+ if (ret && mlx5_removed(priv))
+ return -ENODEV;
return -ret;
}
@@ -282,5 +284,7 @@
mlx5_dev_stop(dev);
mlx5_dev_start(dev);
}
+ if (ret && mlx5_removed(priv))
+ return -ENODEV;
return -ret;
}
@@ -278,6 +278,8 @@
(*priv->rxqs)[idx] = &rxq_ctrl->rxq;
out:
priv_unlock(priv);
+ if (mlx5_removed(priv))
+ return -ENODEV;
return -ret;
}
@@ -485,8 +487,11 @@
}
exit:
priv_unlock(priv);
- if (ret)
+ if (ret) {
WARN("unable to arm interrupt on rx queue %d", rx_queue_id);
+ if (mlx5_removed(priv))
+ return -ENODEV;
+ }
return -ret;
}
@@ -537,9 +542,12 @@
if (rxq_ibv)
mlx5_priv_rxq_ibv_release(priv, rxq_ibv);
priv_unlock(priv);
- if (ret)
+ if (ret) {
WARN("unable to disable interrupt on rx queue %d",
rx_queue_id);
+ if (mlx5_removed(priv))
+ return -ENODEV;
+ }
return -ret;
}
@@ -438,13 +438,17 @@ struct mlx5_counter_ctrl {
stats_n = priv_ethtool_get_stats_n(priv);
if (stats_n < 0) {
priv_unlock(priv);
- return -1;
+ ret = -1;
+ goto error;
}
if (xstats_ctrl->stats_n != stats_n)
priv_xstats_init(priv);
ret = priv_xstats_get(priv, stats);
priv_unlock(priv);
}
+error:
+ if (ret < 0 && mlx5_removed(priv))
+ return -ENODEV;
return ret;
}
@@ -186,6 +186,8 @@
(*priv->txqs)[idx] = &txq_ctrl->txq;
out:
priv_unlock(priv);
+ if (mlx5_removed(priv))
+ return -ENODEV;
return -ret;
}