On Thu, 17 Sep 2026 12:40:08 +0530 Rahul Bhansali <[email protected]> wrote:
> From: Aarnav JP <[email protected]> > > The CPT CQ interrupt handler (nix_inl_cpt_cq_cb) unconditionally > dereferences lf->dev->roc_nix to obtain roc_nix, nix, and port_id. > For inbound, the CPT LF belongs to the inline device which is not > an ethdev, so roc_nix is NULL and the dereference crashes. > > Additionally, error paths returned without writing CPT_LF_DONE_ACK, > leaving CQ entries unacknowledged causing the completion queue to > fill up. > > Fix by deferring roc_nix/nix/port_id derivation into the outbound > branch where roc_nix is valid, setting port_id to UINT32_MAX for > inbound, and routing all error paths through a common cq_ack label > that drains entries and writes CPT_LF_DONE_ACK. > > Fixes: 3fdf3e53f3c4 ("common/cnxk: enable CPT CQ for inline IPsec inbound") > Cc: [email protected] > > Signed-off-by: Aarnav JP <[email protected]> > --- More detailed AI review: Patch 6/14 common/cnxk: fix null deref and irq ack in CPT CQ handler ------------------------------------------------------------------- Info: at cq_ack, head = (lf->cq_head + count) % lf->cq_size; overwrites the head the loop already computed on the normal path. It is only needed for the early gotos.
