On Mon, Sep 15, 2025 at 12:24:16PM +0530, Mahanta Jambigi wrote:

On 12/09/25 2:37 pm, Leon Romanovsky wrote:

quoted

On Fri, Sep 12, 2025 at 01:18:52PM +0530, Mahanta Jambigi wrote:

quoted

On 10/09/25 3:31 pm, Leon Romanovsky wrote:

quoted

quoted

--- a/net/smc/smc_pnet.c
+++ b/net/smc/smc_pnet.c

@@ -450,7 +450,7 @@ static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
 		return -ENOMEM;
 	new_pe->type = SMC_PNET_IB;
 	memcpy(new_pe->pnet_name, pnet_name, SMC_MAX_PNETID_LEN);
-	strncpy(new_pe->ib_name, ib_name, IB_DEVICE_NAME_MAX);
+	strscpy(new_pe->ib_name, ib_name);

It is worth to mention that caching ib_name is wrong as IB/core provides
IB device rename functionality.

In our case we hit this code path where we pass *PCI_ID*
as the *ib_name* using *smc_pnet* tool(smc_pnet -a <pnet_name> -D
<PCI_ID>). I believe PCI_ID will not change, so caching it here is fine.

If I remember, you are reporting that cached ib_name through netlink much later.

The caching itself is not an issue, but incorrect reported name can be seen as
a wrong thing to do.

In what case we can see this incorrect reported name, could you please
elaborate.

Did you open net/smc/smc_pnet.c?

Everything that uses ib_name in that file is incorrect.

From glance look:
1. smc_pnet_find_ib() returns completely random results if device is
renamed in parallel.
2. SMC_PNETID_GET returns wrong names. It returns cached name which
doesn't exist anymore.

IB devices have stable indexes in similar way to netdevice. The code
should rely on it and not on the name.

Thanks