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]>
---
drivers/common/cnxk/roc_nix_inl_dev_irq.c | 48 ++++++++++++++++-------
1 file changed, 33 insertions(+), 15 deletions(-)
diff --git a/drivers/common/cnxk/roc_nix_inl_dev_irq.c
b/drivers/common/cnxk/roc_nix_inl_dev_irq.c
index afbf966f78..f99c32f30b 100644
--- a/drivers/common/cnxk/roc_nix_inl_dev_irq.c
+++ b/drivers/common/cnxk/roc_nix_inl_dev_irq.c
@@ -48,41 +48,57 @@ nix_inl_sso_work_cb(struct nix_inl_dev *inl_dev)
static void
nix_inl_cpt_cq_cb(struct roc_cpt_lf *lf)
{
- struct roc_nix *roc_nix = (struct roc_nix *)lf->dev->roc_nix;
- struct nix *nix = roc_nix_to_nix_priv(roc_nix);
struct idev_cfg *idev = idev_get_cfg();
- uint32_t port_id = roc_nix->port_id;
struct nix_inl_dev *inl_dev = NULL;
enum nix_inl_event_type cq_type;
union cpt_lf_cq_base cq_base;
union cpt_lf_cq_ptr cq_ptr;
+ struct roc_nix *roc_nix;
struct cpt_cq_s *cq_s;
uint8_t fmt_msk = 0x3;
uint32_t count, head;
+ uint32_t port_id = UINT32_MAX;
uint32_t nq_ptr;
+ struct nix *nix;
uint64_t i;
void *sa;
+ /* Read CQ state early so we can always acknowledge the interrupt */
+ head = lf->cq_head;
+ cq_base.u = plt_read64(lf->rbase + CPT_LF_CQ_BASE);
+ cq_ptr.u = plt_read64(lf->rbase + CPT_LF_CQ_PTR);
+ count = cq_ptr.s.count;
+ nq_ptr = cq_ptr.s.nq_ptr;
+
if (idev)
inl_dev = idev->nix_inl_dev;
if (!inl_dev) {
plt_nix_dbg("Inline Device could not be detected");
- return;
+ goto cq_ack;
}
- head = lf->cq_head;
- cq_base.u = plt_read64(lf->rbase + CPT_LF_CQ_BASE);
- cq_ptr.u = plt_read64(lf->rbase + CPT_LF_CQ_PTR);
- count = cq_ptr.s.count;
- nq_ptr = cq_ptr.s.nq_ptr;
-
- if (lf->dev == &inl_dev->dev)
+ if (lf->dev == &inl_dev->dev) {
+ /* Inbound: CPT LF belongs to inline device.
+ * roc_nix is NULL here as inline dev is not an ethdev.
+ * port_id will be derived from SA in the PMD work callback.
+ */
cq_type = NIX_INL_INB_CPT_CQ;
- else if (lf->dev == &nix->dev)
+ } else {
+ /* Outbound: CPT LF belongs to an ethdev */
+ roc_nix = (struct roc_nix *)lf->dev->roc_nix;
+ if (!roc_nix) {
+ plt_nix_dbg("CPT LF dev has no roc_nix");
+ goto cq_ack;
+ }
+ nix = roc_nix_to_nix_priv(roc_nix);
+ if (lf->dev != &nix->dev) {
+ plt_nix_dbg("CPT LF dev mismatch with nix dev");
+ goto cq_ack;
+ }
cq_type = NIX_INL_OUTB_CPT_CQ;
- else
- return;
+ port_id = roc_nix->port_id;
+ }
for (i = 0; i < count; i++) {
cq_s = (struct cpt_cq_s *)(uintptr_t)(((cq_base.s.addr << 7)) +
(head << 5));
@@ -106,11 +122,13 @@ nix_inl_cpt_cq_cb(struct roc_cpt_lf *lf)
head = (head + 1) % lf->cq_size;
}
+cq_ack:
+ /* Drain unprocessed entries and acknowledge the interrupt */
+ head = (lf->cq_head + count) % lf->cq_size;
lf->cq_head = head;
if (unlikely(nq_ptr != head))
plt_err("CPT LF[%d] CQ head %d != NQ ptr %d", lf->lf_id, head,
nq_ptr);
- /* Acknowledge the number of completed requests */
plt_write64(count, lf->rbase + CPT_LF_DONE_ACK);
}
--
2.34.1
