Thanks for dealing with this.
On Thu, Feb 08, 2018 at 04:34:12PM +0000, Matan Azrad wrote:
> Fail-safe PMD uses per sub-device flag called "remove" to indicate the
> scope where the sub-device isn't synchronized with the fail-safe state.
> This flag is set when fail-safe gets RMV notification about the
> physical removal of the sub-device and should be unset when the
> sub-device completes all the configurations cause it to arrive to the
> fail-safe state.
> The previous code wrongly unsets the flag after calling to the
> sub-device PMD dev_configure() operation and before all the
> configurations were done.
> Change the remove flag unsetting to be only after the sub-device
> successes to arrive to the fail-safe state.
I'm not sure this is the right way to do this.
I think it's clear that it was a mistake to set sdev->remove to 0
only during fs_dev_configure.
The flag itself only means "there is something to be done on this
device, please clean up".
Once the clean-up has happened, then the flag is not necessary anymore
and should be reset.
So I thought that this fix would actually put the flag reset within
fs_dev_remove, right before reinstalling the hotplug alarm.
At this point, the device state would have been set back to
DEV_UNDEFINED, so the remove flag is unnecessary for any operation
trying to avoid unplugged slaves.
The "remove" flag is initialized at 0 when sub-devices are allocated
(during fail-safe init). This means that there would be a difference in
the state of the slave between its first initialization and any
subsequent init, after one successful plugout.
> Fixes: a46f8d5 ("net/failsafe: add fail-safe PMD")
> Cc: sta...@dpdk.org
> Signed-off-by: Matan Azrad <ma...@mellanox.com>
> drivers/net/failsafe/failsafe_ether.c | 2 ++
> drivers/net/failsafe/failsafe_ops.c | 2 +-
> 2 files changed, 3 insertions(+), 1 deletion(-)
> diff --git a/drivers/net/failsafe/failsafe_ether.c
> index 4c6e938..ca42376 100644
> --- a/drivers/net/failsafe/failsafe_ether.c
> +++ b/drivers/net/failsafe/failsafe_ether.c
> @@ -377,6 +377,8 @@
> goto err_remove;
> + if (PRIV(dev)->state < DEV_STARTED)
> + sdev->remove = 0;
Here the remove flag should already be 0. If it isn't, this is a
(logical) bug, which should be properly addressed instead of patched
in this way.
> diff --git a/drivers/net/failsafe/failsafe_ops.c
> index 7a67e16..a7c2dba 100644
> --- a/drivers/net/failsafe/failsafe_ops.c
> +++ b/drivers/net/failsafe/failsafe_ops.c
> @@ -131,7 +131,6 @@
> dev->data->dev_conf.intr_conf.lsc = 0;
> DEBUG("Configuring sub-device %d", i);
> - sdev->remove = 0;
This is correct.
> ret = rte_eth_dev_configure(PORT_ID(sdev),
> @@ -197,6 +196,7 @@
> return ret;
> sdev->state = DEV_STARTED;
> + sdev->remove = 0;
This seems unnecessary, if this operation was already performed once the
device has been properly removed.
> if (PRIV(dev)->state < DEV_STARTED)
> PRIV(dev)->state = DEV_STARTED;