]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
ixgbe: fix double clean of Tx descriptors with xdp
authorIlya Maximets <i.maximets@samsung.com>
Thu, 22 Aug 2019 17:12:37 +0000 (20:12 +0300)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Wed, 11 Sep 2019 16:42:18 +0000 (09:42 -0700)
Tx code doesn't clear the descriptors' status after cleaning.
So, if the budget is larger than number of used elems in a ring, some
descriptors will be accounted twice and xsk_umem_complete_tx will move
prod_tail far beyond the prod_head breaking the completion queue ring.

Fix that by limiting the number of descriptors to clean by the number
of used descriptors in the Tx ring.

'ixgbe_clean_xdp_tx_irq()' function refactored to look more like
'ixgbe_xsk_clean_tx_ring()' since we're allowed to directly use
'next_to_clean' and 'next_to_use' indexes.

CC: stable@vger.kernel.org
Fixes: 8221c5eba8c1 ("ixgbe: add AF_XDP zero-copy Tx support")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Tested-by: William Tu <u9012063@gmail.com>
Tested-by: Eelco Chaudron <echaudro@redhat.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/ixgbe/ixgbe_xsk.c

index 6b609553329fa98647b78f4e1e5d3e9223a8aa7a..a3b6d8c89127f1a2fcfdfb84a1824dbbedc019e6 100644 (file)
@@ -633,19 +633,17 @@ static void ixgbe_clean_xdp_tx_buffer(struct ixgbe_ring *tx_ring,
 bool ixgbe_clean_xdp_tx_irq(struct ixgbe_q_vector *q_vector,
                            struct ixgbe_ring *tx_ring, int napi_budget)
 {
+       u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
        unsigned int total_packets = 0, total_bytes = 0;
-       u32 i = tx_ring->next_to_clean, xsk_frames = 0;
-       unsigned int budget = q_vector->tx.work_limit;
        struct xdp_umem *umem = tx_ring->xsk_umem;
        union ixgbe_adv_tx_desc *tx_desc;
        struct ixgbe_tx_buffer *tx_bi;
-       bool xmit_done;
+       u32 xsk_frames = 0;
 
-       tx_bi = &tx_ring->tx_buffer_info[i];
-       tx_desc = IXGBE_TX_DESC(tx_ring, i);
-       i -= tx_ring->count;
+       tx_bi = &tx_ring->tx_buffer_info[ntc];
+       tx_desc = IXGBE_TX_DESC(tx_ring, ntc);
 
-       do {
+       while (ntc != ntu) {
                if (!(tx_desc->wb.status & cpu_to_le32(IXGBE_TXD_STAT_DD)))
                        break;
 
@@ -661,22 +659,18 @@ bool ixgbe_clean_xdp_tx_irq(struct ixgbe_q_vector *q_vector,
 
                tx_bi++;
                tx_desc++;
-               i++;
-               if (unlikely(!i)) {
-                       i -= tx_ring->count;
+               ntc++;
+               if (unlikely(ntc == tx_ring->count)) {
+                       ntc = 0;
                        tx_bi = tx_ring->tx_buffer_info;
                        tx_desc = IXGBE_TX_DESC(tx_ring, 0);
                }
 
                /* issue prefetch for next Tx descriptor */
                prefetch(tx_desc);
+       }
 
-               /* update budget accounting */
-               budget--;
-       } while (likely(budget));
-
-       i += tx_ring->count;
-       tx_ring->next_to_clean = i;
+       tx_ring->next_to_clean = ntc;
 
        u64_stats_update_begin(&tx_ring->syncp);
        tx_ring->stats.bytes += total_bytes;
@@ -688,8 +682,7 @@ bool ixgbe_clean_xdp_tx_irq(struct ixgbe_q_vector *q_vector,
        if (xsk_frames)
                xsk_umem_complete_tx(umem, xsk_frames);
 
-       xmit_done = ixgbe_xmit_zc(tx_ring, q_vector->tx.work_limit);
-       return budget > 0 && xmit_done;
+       return ixgbe_xmit_zc(tx_ring, q_vector->tx.work_limit);
 }
 
 int ixgbe_xsk_async_xmit(struct net_device *dev, u32 qid)