From: Ido Schimmel idosch@nvidia.com
stable inclusion from stable-5.10.77 commit e81bed557fe7dedff78ac1a7c098f4375b4981e2 bugzilla: 185677 https://gitee.com/openeuler/kernel/issues/I4IAP7
Reference: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id=...
--------------------------------
commit 759635760a804b0d8ad0cc677b650f1544cae22f upstream.
When the driver fails to allocate a new Rx buffer, it passes an empty Rx descriptor (contains zero address and size) to the device and marks it as invalid by setting the skb pointer in the descriptor's metadata to NULL.
After processing enough Rx descriptors, the driver will try to process the invalid descriptor, but will return immediately seeing that the skb pointer is NULL. Since the driver no longer passes new Rx descriptors to the device, the Rx queue will eventually become full and the device will start to drop packets.
Fix this by recycling the received packet if allocation of the new packet failed. This means that allocation is no longer performed at the end of the Rx routine, but at the start, before tearing down the DMA mapping of the received packet.
Remove the comment about the descriptor being zeroed as it is no longer correct. This is OK because we either use the descriptor as-is (when recycling) or overwrite its address and size fields with that of the newly allocated Rx buffer.
The issue was discovered when a process ("perf") consumed too much memory and put the system under memory pressure. It can be reproduced by injecting slab allocation failures [1]. After the fix, the Rx queue no longer comes to a halt.
[1] # echo 10 > /sys/kernel/debug/failslab/times # echo 1000 > /sys/kernel/debug/failslab/interval # echo 100 > /sys/kernel/debug/failslab/probability
FAULT_INJECTION: forcing a failure. name failslab, interval 1000, probability 100, space 0, times 8 [...] Call Trace: <IRQ> dump_stack_lvl+0x34/0x44 should_fail.cold+0x32/0x37 should_failslab+0x5/0x10 kmem_cache_alloc_node+0x23/0x190 __alloc_skb+0x1f9/0x280 __netdev_alloc_skb+0x3a/0x150 mlxsw_pci_rdq_skb_alloc+0x24/0x90 mlxsw_pci_cq_tasklet+0x3dc/0x1200 tasklet_action_common.constprop.0+0x9f/0x100 __do_softirq+0xb5/0x252 irq_exit_rcu+0x7a/0xa0 common_interrupt+0x83/0xa0 </IRQ> asm_common_interrupt+0x1e/0x40 RIP: 0010:cpuidle_enter_state+0xc8/0x340 [...] mlxsw_spectrum2 0000:06:00.0: Failed to alloc skb for RDQ
Fixes: eda6500a987a ("mlxsw: Add PCI bus implementation") Signed-off-by: Ido Schimmel idosch@nvidia.com Reviewed-by: Petr Machata petrm@nvidia.com Link: https://lore.kernel.org/r/20211024064014.1060919-1-idosch@idosch.org Signed-off-by: Jakub Kicinski kuba@kernel.org Signed-off-by: Greg Kroah-Hartman gregkh@linuxfoundation.org Signed-off-by: Chen Jun chenjun102@huawei.com Acked-by: Weilong Chen chenweilong@huawei.com
Signed-off-by: Chen Jun chenjun102@huawei.com --- drivers/net/ethernet/mellanox/mlxsw/pci.c | 25 +++++++++++------------ 1 file changed, 12 insertions(+), 13 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlxsw/pci.c b/drivers/net/ethernet/mellanox/mlxsw/pci.c index 641cdd81882b..ffaeda75eec4 100644 --- a/drivers/net/ethernet/mellanox/mlxsw/pci.c +++ b/drivers/net/ethernet/mellanox/mlxsw/pci.c @@ -353,13 +353,10 @@ static int mlxsw_pci_rdq_skb_alloc(struct mlxsw_pci *mlxsw_pci, struct sk_buff *skb; int err;
- elem_info->u.rdq.skb = NULL; skb = netdev_alloc_skb_ip_align(NULL, buf_len); if (!skb) return -ENOMEM;
- /* Assume that wqe was previously zeroed. */ - err = mlxsw_pci_wqe_frag_map(mlxsw_pci, wqe, 0, skb->data, buf_len, DMA_FROM_DEVICE); if (err) @@ -548,21 +545,26 @@ static void mlxsw_pci_cqe_rdq_handle(struct mlxsw_pci *mlxsw_pci, struct pci_dev *pdev = mlxsw_pci->pdev; struct mlxsw_pci_queue_elem_info *elem_info; struct mlxsw_rx_info rx_info = {}; - char *wqe; + char wqe[MLXSW_PCI_WQE_SIZE]; struct sk_buff *skb; u16 byte_count; int err;
elem_info = mlxsw_pci_queue_elem_info_consumer_get(q); - skb = elem_info->u.sdq.skb; - if (!skb) - return; - wqe = elem_info->elem; - mlxsw_pci_wqe_frag_unmap(mlxsw_pci, wqe, 0, DMA_FROM_DEVICE); + skb = elem_info->u.rdq.skb; + memcpy(wqe, elem_info->elem, MLXSW_PCI_WQE_SIZE);
if (q->consumer_counter++ != consumer_counter_limit) dev_dbg_ratelimited(&pdev->dev, "Consumer counter does not match limit in RDQ\n");
+ err = mlxsw_pci_rdq_skb_alloc(mlxsw_pci, elem_info); + if (err) { + dev_err_ratelimited(&pdev->dev, "Failed to alloc skb for RDQ\n"); + goto out; + } + + mlxsw_pci_wqe_frag_unmap(mlxsw_pci, wqe, 0, DMA_FROM_DEVICE); + if (mlxsw_pci_cqe_lag_get(cqe_v, cqe)) { rx_info.is_lag = true; rx_info.u.lag_id = mlxsw_pci_cqe_lag_id_get(cqe_v, cqe); @@ -594,10 +596,7 @@ static void mlxsw_pci_cqe_rdq_handle(struct mlxsw_pci *mlxsw_pci, skb_put(skb, byte_count); mlxsw_core_skb_receive(mlxsw_pci->core, skb, &rx_info);
- memset(wqe, 0, q->elem_size); - err = mlxsw_pci_rdq_skb_alloc(mlxsw_pci, elem_info); - if (err) - dev_dbg_ratelimited(&pdev->dev, "Failed to alloc skb for RDQ\n"); +out: /* Everything is set up, ring doorbell to pass elem to HW */ q->producer_counter++; mlxsw_pci_queue_doorbell_producer_ring(mlxsw_pci, q);