[dpdk-dev] [PATCH] net/bonding: fix buf corruption in merging un-transmitted packets

Jia Yu jyu at vmware.com
Mon Aug 20 09:02:44 CEST 2018


Hi Chas,

Could you help to review the v2 of the patch? I have addressed the “Fixes” requirement in commit message, and modified slave_tx_fail_count’s definition.

Thanks,
Jia

From: Chas Williams <3chas3 at gmail.com>
Date: Sunday, August 19, 2018 at 5:07 PM
To: Jia Yu <jyu at vmware.com>
Cc: "dev at dpdk.org" <dev at dpdk.org>, Declan Doherty <declan.doherty at intel.com>, Chas Williams <chas3 at att.com>
Subject: Re: [dpdk-dev] [PATCH] net/bonding: fix buf corruption in merging un-transmitted packets



On Sun, Aug 19, 2018 at 6:19 PM Jia Yu <jyu at vmware.com<mailto:jyu at vmware.com>> wrote:
Hi Chas,

Thanks for reviewing the change.

Our application crashed after upgrading to DPDK 18.02, when packet rate is high and bond is configured. It happened because txq contains invalid mbuf addresses after rte_eth_tx_burst call (for example, 0x100000000 repeated 13 times in one core dump). It seems that the invalid addresses came from buf shift code below, so I changed this part of code to earlier version. We don’t see crash after the fix. Please let us know if the fix is reasonable.

With respect to correctness, the code you are fixing does seem broken.
See inline below.


m_table = {0x7fdf23a68ec0, 0x7fdf23a68400, 0x7fdf23a66e80, 0x7fdf23a65900,
    0x7fdf23960e00, 0x100000000 <repeats 13 times>, 0x7fdf23978640, 0x7fdf23977b80, 0x7fdf239745c0, 0x7fdf23a4d600, 0x7fdf23a4cb40,
    0x7fdf23a75b00, 0x7fdf23a74580, 0x7fdf23972580, 0x7fdf239a76c0, 0x7fdf239a5680, 0x7fdf23a7a640, 0x7fdf23a79b80, 0x7fdf23a77b40,
    0x7fdf2395b800}

/* Send packet burst on each slave device */
    for (i = 0; i < slave_count; i++) {
        if (slave_nb_bufs[i] == 0)
            continue;

        slave_tx_count = rte_eth_tx_burst(slave_port_ids[i],
                bd_tx_q->queue_id, slave_bufs[i],
                slave_nb_bufs[i]);

A typical failure here would be to transmit no packets.  slave_tx_count = 0.




        total_tx_count += slave_tx_count;

        /* If tx burst fails move packets to end of bufs */
        if (unlikely(slave_tx_count < slave_nb_bufs[i])) {
            slave_tx_fail_count[i] = slave_nb_bufs[i] -
                    slave_tx_count;
            total_tx_fail_count += slave_tx_fail_count[i];

            /*
             * Shift bufs to beginning of array to allow reordering
             * later
             */
            for (j = 0; j < slave_tx_fail_count[i]; j++) {
                slave_bufs[i][j] =
                    slave_bufs[i][(slave_tx_count - 1) + j]; <==== what if slave_tx_count == 0, j == 0

As you correctly point out, slave_tx_count - 1 would be -1.  Bad news.
So, yes, I agree with your fix.  In your fix, you might notice that
slave_tx_fail_count doesn't need to be an array.  It's local to
if (unlikely(slave_tx_count < slave_nb_bufs[i])) now.


            }
        }
    }

Thanks,
Jia

From: Chas Williams <3chas3 at gmail.com<mailto:3chas3 at gmail.com>>
Date: Saturday, August 18, 2018 at 4:50 PM
To: Jia Yu <jyu at vmware.com<mailto:jyu at vmware.com>>
Cc: "dev at dpdk.org<mailto:dev at dpdk.org>" <dev at dpdk.org<mailto:dev at dpdk.org>>, Declan Doherty <declan.doherty at intel.com<mailto:declan.doherty at intel.com>>, Chas Williams <chas3 at att.com<mailto:chas3 at att.com>>
Subject: Re: [dpdk-dev] [PATCH] net/bonding: fix buf corruption in merging un-transmitted packets



On Fri, Aug 17, 2018 at 9:46 PM Jia Yu <jyu at vmware.com<mailto:jyu at vmware.com>> wrote:
When bond slave devices cannot transmit all packets in bufs array,
tx_burst callback shall merge the un-transmitted packets back to
bufs array. Recent merge logic introduced a bug which causes
invalid mbuf addresses being written to bufs array.

Can you expand on this a bit?  What was the commit?


When caller frees the un-transmitted packets, due to invalid addresses,
application will crash.

The fix is avoid shifting mbufs, and directly write un-transmitted
packets back to bufs array.

Signed-off-by: Jia Yu <jyu at vmware.com<mailto:jyu at vmware.com>>
---
 drivers/net/bonding/rte_eth_bond_pmd.c | 98 +++++-----------------------------
 1 file changed, 14 insertions(+), 84 deletions(-)

diff --git a/drivers/net/bonding/rte_eth_bond_pmd.c b/drivers/net/bonding/rte_eth_bond_pmd.c
index 58f7377..cce973a 100644
--- a/drivers/net/bonding/rte_eth_bond_pmd.c
+++ b/drivers/net/bonding/rte_eth_bond_pmd.c
@@ -303,7 +303,7 @@ bond_ethdev_tx_burst_8023ad_fast_queue(void *queue, struct rte_mbuf **bufs,
        uint16_t slave_tx_count, slave_tx_fail_count[RTE_MAX_ETHPORTS] = { 0 };
        uint16_t total_tx_count = 0, total_tx_fail_count = 0;

-       uint16_t i, j;
+       uint16_t i;

        if (unlikely(nb_bufs == 0))
                return 0;
@@ -361,31 +361,9 @@ bond_ethdev_tx_burst_8023ad_fast_queue(void *queue, struct rte_mbuf **bufs,
                        slave_tx_fail_count[i] = slave_nb_bufs[i] -
                                        slave_tx_count;
                        total_tx_fail_count += slave_tx_fail_count[i];
-
-                       /*
-                        * Shift bufs to beginning of array to allow reordering
-                        * later
-                        */
-                       for (j = 0; j < slave_tx_fail_count[i]; j++) {
-                               slave_bufs[i][j] =
-                                       slave_bufs[i][(slave_tx_count - 1) + j];
-                       }
-               }
-       }
-
-       /*
-        * If there are tx burst failures we move packets to end of bufs to
-        * preserve expected PMD behaviour of all failed transmitted being
-        * at the end of the input mbuf array
-        */
-       if (unlikely(total_tx_fail_count > 0)) {
-               int bufs_idx = nb_bufs - total_tx_fail_count - 1;
-
-               for (i = 0; i < slave_count; i++) {
-                       if (slave_tx_fail_count[i] > 0) {
-                               for (j = 0; j < slave_tx_fail_count[i]; j++)
-                                       bufs[bufs_idx++] = slave_bufs[i][j];
-                       }
+                       memcpy(&bufs[nb_bufs - total_tx_fail_count],
+                              &slave_bufs[i][slave_tx_count],
+                              slave_tx_fail_count[i] * sizeof(bufs[0]));
                }
        }

@@ -715,8 +693,8 @@ bond_ethdev_tx_burst_round_robin(void *queue, struct rte_mbuf **bufs,
                                tx_fail_total += tx_fail_slave;

                                memcpy(&bufs[nb_pkts - tx_fail_total],
-                                               &slave_bufs[i][num_tx_slave],
-                                               tx_fail_slave * sizeof(bufs[0]));
+                                      &slave_bufs[i][num_tx_slave],
+                                      tx_fail_slave * sizeof(bufs[0]));
                        }
                        num_tx_total += num_tx_slave;
                }
@@ -1224,7 +1202,7 @@ bond_ethdev_tx_burst_balance(void *queue, struct rte_mbuf **bufs,
        uint16_t slave_tx_count, slave_tx_fail_count[RTE_MAX_ETHPORTS] = { 0 };
        uint16_t total_tx_count = 0, total_tx_fail_count = 0;

-       uint16_t i, j;
+       uint16_t i;

        if (unlikely(nb_bufs == 0))
                return 0;
@@ -1268,31 +1246,9 @@ bond_ethdev_tx_burst_balance(void *queue, struct rte_mbuf **bufs,
                        slave_tx_fail_count[i] = slave_nb_bufs[i] -
                                        slave_tx_count;
                        total_tx_fail_count += slave_tx_fail_count[i];
-
-                       /*
-                        * Shift bufs to beginning of array to allow reordering
-                        * later
-                        */
-                       for (j = 0; j < slave_tx_fail_count[i]; j++) {
-                               slave_bufs[i][j] =
-                                       slave_bufs[i][(slave_tx_count - 1) + j];
-                       }
-               }
-       }
-
-       /*
-        * If there are tx burst failures we move packets to end of bufs to
-        * preserve expected PMD behaviour of all failed transmitted being
-        * at the end of the input mbuf array
-        */
-       if (unlikely(total_tx_fail_count > 0)) {
-               int bufs_idx = nb_bufs - total_tx_fail_count - 1;
-
-               for (i = 0; i < slave_count; i++) {
-                       if (slave_tx_fail_count[i] > 0) {
-                               for (j = 0; j < slave_tx_fail_count[i]; j++)
-                                       bufs[bufs_idx++] = slave_bufs[i][j];
-                       }
+                       memcpy(&bufs[nb_bufs - total_tx_fail_count],
+                              &slave_bufs[i][slave_tx_count],
+                              slave_tx_fail_count[i] * sizeof(bufs[0]));
                }
        }

@@ -1322,7 +1278,7 @@ bond_ethdev_tx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
        uint16_t slave_tx_count, slave_tx_fail_count[RTE_MAX_ETHPORTS] = { 0 };
        uint16_t total_tx_count = 0, total_tx_fail_count = 0;

-       uint16_t i, j;
+       uint16_t i;

        if (unlikely(nb_bufs == 0))
                return 0;
@@ -1384,35 +1340,9 @@ bond_ethdev_tx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
                                                slave_tx_count;
                                total_tx_fail_count += slave_tx_fail_count[i];

-                               /*
-                                * Shift bufs to beginning of array to allow
-                                * reordering later
-                                */
-                               for (j = 0; j < slave_tx_fail_count[i]; j++)
-                                       slave_bufs[i][j] =
-                                               slave_bufs[i]
-                                                       [(slave_tx_count - 1)
-                                                       + j];
-                       }
-               }
-
-               /*
-                * If there are tx burst failures we move packets to end of
-                * bufs to preserve expected PMD behaviour of all failed
-                * transmitted being at the end of the input mbuf array
-                */
-               if (unlikely(total_tx_fail_count > 0)) {
-                       int bufs_idx = nb_bufs - total_tx_fail_count - 1;
-
-                       for (i = 0; i < slave_count; i++) {
-                               if (slave_tx_fail_count[i] > 0) {
-                                       for (j = 0;
-                                               j < slave_tx_fail_count[i];
-                                               j++) {
-                                               bufs[bufs_idx++] =
-                                                       slave_bufs[i][j];
-                                       }
-                               }
+                               memcpy(&bufs[nb_bufs - total_tx_fail_count],
+                                      &slave_bufs[i][slave_tx_count],
+                                      slave_tx_fail_count[i] * sizeof(bufs[0]));
                        }
                }
        }
--
2.7.4


More information about the dev mailing list