From 25c01bd32374b0c3cbc260f3e3872408d749cb45 Mon Sep 17 00:00:00 2001
From: Matan Azrad <matan@nvidia.com>
Date: Thu, 11 Aug 2022 19:59:18 +0300
Subject: [PATCH] net/mlx5: fix Rx queue recovery mechanism

[ upstream commit 60b254e3923d007bcadbb8d410f95ad89a2f13fa ]

The local variables are getting inconsistent in data receiving routines
after queue error recovery.
Receive queue consumer index is getting wrong, need to reset one to the
size of the queue (as RQ was fully replenished in recovery procedure).

In MPRQ case, also the local consumed strd variable should be reset.

CVE-2022-28199
Fixes: 88c0733535d6 ("net/mlx5: extend Rx completion with error handling")

Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Signed-off-by: Matan Azrad <matan@nvidia.com>
Conflict: NA
Reference: https://git.dpdk.org/dpdk-stable/commit/?id=25c01db32374
---
 drivers/net/mlx5/mlx5_rx.c | 34 ++++++++++++++++++++++++----------
 1 file changed, 24 insertions(+), 10 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_rx.c b/drivers/net/mlx5/mlx5_rx.c
index f388fcc313..9fcd039c22 100644
--- a/drivers/net/mlx5/mlx5_rx.c
+++ b/drivers/net/mlx5/mlx5_rx.c
@@ -390,6 +390,11 @@ mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
 }
 
+/* Must be negative. */
+#define MLX5_ERROR_CQE_RET (-1)
+/* Must not be negative. */
+#define MLX5_RECOVERY_ERROR_RET 0
+
 /**
  * Handle a Rx error.
  * The function inserts the RQ state to reset when the first error CQE is
@@ -404,7 +409,7 @@ mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
  *   0 when called from non-vectorized Rx burst.
  *
  * @return
- *   -1 in case of recovery error, otherwise the CQE status.
+ *   MLX5_RECOVERY_ERROR_RET in case of recovery error, otherwise the CQE status.
  */
 int
 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
@@ -433,7 +438,7 @@ mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
 		sm.queue_id = rxq->idx;
 		sm.state = IBV_WQS_RESET;
 		if (mlx5_queue_state_modify(RXQ_DEV(rxq_ctrl), &sm))
-			return -1;
+			return MLX5_RECOVERY_ERROR_RET;
 		if (rxq_ctrl->dump_file_n <
 		    RXQ_PORT(rxq_ctrl)->config.max_dump_files_num) {
 			MKSTR(err_str, "Unexpected CQE error syndrome "
@@ -473,7 +478,7 @@ mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
 			sm.queue_id = rxq->idx;
 			sm.state = IBV_WQS_RDY;
 			if (mlx5_queue_state_modify(RXQ_DEV(rxq_ctrl), &sm))
-				return -1;
+				return MLX5_RECOVERY_ERROR_RET;
 			if (vec) {
 				const uint32_t elts_n =
 					mlx5_rxq_mprq_enabled(rxq) ?
@@ -501,7 +506,7 @@ mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
 							rte_pktmbuf_free_seg
 								(*elt);
 						}
-						return -1;
+						return MLX5_RECOVERY_ERROR_RET;
 					}
 				}
 				for (i = 0; i < (int)elts_n; ++i) {
@@ -520,7 +525,7 @@ mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
 		}
 		return ret;
 	default:
-		return -1;
+		return MLX5_RECOVERY_ERROR_RET;
 	}
 }
 
@@ -538,7 +543,9 @@ mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
  *   written.
  *
  * @return
- *   0 in case of empty CQE, otherwise the packet size in bytes.
+ *   0 in case of empty CQE, MLX5_ERROR_CQE_RET in case of error CQE,
+ *   otherwise the packet size in regular RxQ, and striding byte
+ *   count format in mprq case.
  */
 static inline int
 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
@@ -605,8 +612,8 @@ mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
 					     rxq->err_state)) {
 					ret = mlx5_rx_err_handle(rxq, 0);
 					if (ret == MLX5_CQE_STATUS_HW_OWN ||
-					    ret == -1)
-						return 0;
+					    ret == MLX5_RECOVERY_ERROR_RET)
+						return MLX5_ERROR_CQE_RET;
 				} else {
 					return 0;
 				}
@@ -851,8 +858,10 @@ mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		if (!pkt) {
 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
-			if (!len) {
+			if (len <= 0) {
 				rte_mbuf_raw_free(rep);
+				if (unlikely(len == MLX5_ERROR_CQE_RET))
+					rq_ci = rxq->rq_ci << sges_n;
 				break;
 			}
 			pkt = seg;
@@ -1075,8 +1084,13 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		}
 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
-		if (!ret)
+		if (ret == 0)
+			break;
+		if (unlikely(ret == MLX5_ERROR_CQE_RET)) {
+			rq_ci = rxq->rq_ci;
+			consumed_strd = rxq->consumed_strd;
 			break;
+		}
 		byte_cnt = ret;
 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
 		MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
-- 
2.23.0