From: Bryam Vargas smcr_lgr_reg_rmbs() skips the MR registration (its fast path) when the RMB is registered on every active link, but the check reads is_reg_mr[] at the index of the connection's link for each link it walks. An RMB that sits unused in the pool when a link is added is not registered on that link. A later connection on the older link that reuses it takes the fast path, never registers it on the newer link, and its CONFIRM_RKEY still gives the peer an rkey for that link. After a failover to the newer link, the peer's RDMA writes into the RMB fail with a remote access error, that link goes down too and, with no other link left, the link group is terminated. Read is_reg_mr[] at each active link's own index, so the fast path is taken only when the RMB is registered on all of them. Fixes: 4da687448de7 ("net/smc: reduce unnecessary blocking in smcr_lgr_reg_rmbs()") Cc: stable@vger.kernel.org Signed-off-by: Bryam Vargas --- Repro on SMC-Rv2 with soft-RoCE (rxe) over two veth pairs between two namespaces. The link group starts with one link, and the second link comes up only after the first connection is closed and its RMBs are back in the pools. A second connection on the first link then takes those RMBs, and the server's rxe device for the first link is removed. Unpatched, the client's first RDMA write over the second link completes with IB_WC_REM_ACCESS_ERR, the second link goes down and the link groups on both sides are terminated; the server reads EOF. Patched, the connection survives the failover and the data checks out in both directions. A connection that gets a new RMB survives the failover with and without the patch, and one placed on the second link that reuses the RMBs takes the slow path in both builds (read from debug counters added to smc.ko for the test). No KASAN splat in any run. --- net/smc/af_smc.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c index bfc0c7e71a12..53ee0f4afc11 100644 --- a/net/smc/af_smc.c +++ b/net/smc/af_smc.c @@ -581,7 +581,7 @@ static int smcr_lgr_reg_rmbs(struct smc_link *link, for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { if (!smc_link_active(&lgr->lnk[i])) continue; - if (!rmb_desc->is_reg_mr[link->link_idx]) { + if (!rmb_desc->is_reg_mr[lgr->lnk[i].link_idx]) { up_read(&lgr->llc_conf_mutex); goto slow_path; } --- base-commit: ead6fa9a0ab194f02a2f36d634b193e9c5f9a1c6 change-id: 20261113-b4-disp-08835f86-e769a35d475b Best regards, -- Bryam Vargas