Thank you for the clarification. 

On Sun, Sep 11, 2022 at 3:18 AM Asaf Penso <asafp@nvidia.com> wrote:

Hello Amiya,

 

This fix is for an issue with the error recovery mechanism.

 

I assume the use case you’re referring to is working with rx_vec burst + cqe zipping enabled + error recovery.

If that’s the case, we mentioned in our mlx5.rst that the recovery is not supported.

Since there is no way to discable the recovery in our PMD, we suggested to disable cqe zipping.

 

However, these days, we are working on a patch to allow the above combination.

It should be sent to the ML in a week or two, once we fully finish testing it.

 

Regards,

Asaf Penso

 

From: Amiya Mohakud <amohakud@paloaltonetworks.com>
Sent: Wednesday, September 7, 2022 9:31 AM
To: NBU-Contact-Thomas Monjalon (EXTERNAL) <thomas@monjalon.net>
Cc: dev <dev@dpdk.org>; security@dpdk.org; Matan Azrad <matan@nvidia.com>; stable@dpdk.org; Alexander Kozyrev <akozyrev@nvidia.com>; Slava Ovsiienko <viacheslavo@nvidia.com>; Shahaf Shuler <shahafs@nvidia.com>
Subject: Re: [PATCH] net/mlx5: fix Rx queue recovery mechanism

 

Hi All,

 

I would need some confirmation on this patch. 

For some earlier issues encountered on mlx5, we have disable cqe_comp in the mlx5 driver. In that case, do we still need this fix or disabling cqe_comp will take care of it as well?

 

Regards

Amiya

 

On Mon, Aug 29, 2022 at 8:45 PM Thomas Monjalon <thomas@monjalon.net> wrote:

From: Matan Azrad <matan@nvidia.com>

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")
Cc: stable@dpdk.org

Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Signed-off-by: Matan Azrad <matan@nvidia.com>
---

Already applied in main branch as part of the public disclosure process.

---
 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 bb3ccc36e5..917c517b83 100644
--- a/drivers/net/mlx5/mlx5_rx.c
+++ b/drivers/net/mlx5/mlx5_rx.c
@@ -408,6 +408,11 @@ 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
@@ -422,7 +427,7 @@ 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)
@@ -451,7 +456,7 @@ 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 "
@@ -491,7 +496,7 @@ 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) ?
@@ -519,7 +524,7 @@ 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) {
@@ -538,7 +543,7 @@ mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
                }
                return ret;
        default:
-               return -1;
+               return MLX5_RECOVERY_ERROR_RET;
        }
 }

@@ -556,7 +561,9 @@ 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,
@@ -623,8 +630,8 @@ 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;
                                }
@@ -869,8 +876,10 @@ 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;
@@ -1093,8 +1102,13 @@ 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));
--
2.36.1