From: Vlad Dumitrescu vdumitrescu@nvidia.com
[ Upstream commit 42f993d3439827c4959ea77e60620d7ebfb3a477 ]
Currently, all master upper netdevices (e.g., bond, VRF) are treated equally.
When a VRF netdevice is used over an IPoIB netdevice, the expected netdev resolution is on the lower IPoIB device which has the IP address assigned to it and not the VRF device.
The rdma_cm module (CMA) tries to match incoming requests to a particular netdevice. When successful, it also validates that the return path points to the same device by performing a routing table lookup. Currently, the former would resolve to the VRF netdevice, while the latter to the correct lower IPoIB netdevice, leading to failure in rdma_cm.
Improve this by ignoring the VRF master netdevice, if it exists, and instead return the lower IPoIB device.
Signed-off-by: Vlad Dumitrescu vdumitrescu@nvidia.com Reviewed-by: Parav Pandit parav@nvidia.com Signed-off-by: Edward Srouji edwards@nvidia.com Link: https://patch.msgid.link/20250916111103.84069-5-edwards@nvidia.com Signed-off-by: Leon Romanovsky leon@kernel.org Signed-off-by: Sasha Levin sashal@kernel.org ---
LLM Generated explanations, may be completely bogus:
YES – this change fixes a real regression for VRF users with almost no risk.
- `ipoib_get_master_net_dev()` now skips L3 masters and returns the underlying IPoIB device (`drivers/infiniband/ulp/ipoib/ipoib_main.c:361-374`). Without this, the helper hands `rdma_cm` the VRF device when matching by pkey/GID; the follow-up route validation in `validate_net_dev()` insists on the same device that the fib lookup returns (`drivers/infiniband/core/cma.c:1589-1616`), so requests fail today. - With the new `netif_is_l3_master()` guard we still return true L2 masters such as bonds, so existing bonding setups stay intact while VRF stacks finally resolve to the only device that actually carries the IP address (same file, same lines). - This helper is static and only called via `ipoib_match_gid_pkey_addr()` (`drivers/infiniband/ulp/ipoib/ipoib_main.c:500-505`), so the fix is tightly scoped; the extra comment edits (`drivers/infiniband/ulp/ipoib/ipoib_main.c:526-541`) are clarifications only. - The buggy behavior has been present since IPoIB added the connection- parameter matching helper (`ddde896e561a5`), so all maintained stable kernels inherit the failure. The fix relies only on long-existing helpers and keeps the same refcounting pattern, making the backport straightforward.
Next step: 1) Validate with an RDMA connection over an IPoIB device enslaved to a VRF to confirm the CMA path succeeds after backport.
drivers/infiniband/ulp/ipoib/ipoib_main.c | 21 +++++++++++---------- 1 file changed, 11 insertions(+), 10 deletions(-)
diff --git a/drivers/infiniband/ulp/ipoib/ipoib_main.c b/drivers/infiniband/ulp/ipoib/ipoib_main.c index 7acafc5c0e09a..5b4d76e97437d 100644 --- a/drivers/infiniband/ulp/ipoib/ipoib_main.c +++ b/drivers/infiniband/ulp/ipoib/ipoib_main.c @@ -351,26 +351,27 @@ static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr, }
/* - * Find the master net_device on top of the given net_device. + * Find the L2 master net_device on top of the given net_device. * @dev: base IPoIB net_device * - * Returns the master net_device with a reference held, or the same net_device - * if no master exists. + * Returns the L2 master net_device with reference held if the L2 master + * exists (such as bond netdevice), or returns same netdev with reference + * held when master does not exist or when L3 master (such as VRF netdev). */ static struct net_device *ipoib_get_master_net_dev(struct net_device *dev) { struct net_device *master;
rcu_read_lock(); + master = netdev_master_upper_dev_get_rcu(dev); + if (!master || netif_is_l3_master(master)) + master = dev; + dev_hold(master); rcu_read_unlock();
- if (master) - return master; - - dev_hold(dev); - return dev; + return master; }
struct ipoib_walk_data { @@ -522,7 +523,7 @@ static struct net_device *ipoib_get_net_dev_by_params( if (ret) return NULL;
- /* See if we can find a unique device matching the L2 parameters */ + /* See if we can find a unique device matching the pkey and GID */ matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index, gid, NULL, &net_dev);
@@ -535,7 +536,7 @@ static struct net_device *ipoib_get_net_dev_by_params(
dev_put(net_dev);
- /* Couldn't find a unique device with L2 parameters only. Use L3 + /* Couldn't find a unique device with pkey and GID only. Use L3 * address to uniquely match the net device */ matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index, gid, addr, &net_dev);