When build_sec_fd fails at index loop inside the enqueue burst loops,
the previously built FD entries (indices 0..loop-1) were never freed.
The cleanup loop iterated in the wrong direction, starting at the
failed index and going up to frames_to_send, which are entries that
were never built. This caused silent FLE pool exhaustion, after which
every subsequent build_sec_fd returned -ENOMEM, enqueue_burst
returned 0 indefinitely, and the crypto-perf test hung.
Fix both dpaa2_sec_enqueue_burst and dpaa2_sec_enqueue_burst_ordered
by clamping frames_to_send to loop + 1 and iterating from 0 to
free all allocated FLE buffers including the failed entry.
Fixes: 623326dded ("crypto/dpaa2_sec: introduce poll mode driver")
Cc: [email protected]
Signed-off-by: Gagandeep Singh <[email protected]>
---
drivers/crypto/dpaa2_sec/dpaa2_sec_dpseci.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/drivers/crypto/dpaa2_sec/dpaa2_sec_dpseci.c
b/drivers/crypto/dpaa2_sec/dpaa2_sec_dpseci.c
index 2a015a3d82..15152cc5a1 100644
--- a/drivers/crypto/dpaa2_sec/dpaa2_sec_dpseci.c
+++ b/drivers/crypto/dpaa2_sec/dpaa2_sec_dpseci.c
@@ -1550,6 +1550,9 @@ dpaa2_sec_enqueue_burst(void *qp, struct rte_crypto_op
**ops,
ret = build_sec_fd(*ops, &fd_arr[loop], bpid, dpaa2_qp);
if (ret) {
DPAA2_SEC_DP_DEBUG("FD build failed");
+ frames_to_send = loop + 1;
+ for (loop = 0; loop < frames_to_send; loop++)
+ free_fle(&fd_arr[loop], dpaa2_qp);
goto skip_tx;
}
ops++;
@@ -1909,6 +1912,9 @@ dpaa2_sec_enqueue_burst_ordered(void *qp, struct
rte_crypto_op **ops,
ret = build_sec_fd(*ops, &fd_arr[loop], bpid, dpaa2_qp);
if (ret) {
DPAA2_SEC_DP_DEBUG("FD build failed");
+ frames_to_send = loop + 1;
+ for (loop = 0; loop < frames_to_send; loop++)
+ free_fle(&fd_arr[loop], dpaa2_qp);
goto skip_tx;
}
ops++;
--
2.25.1
