Commit 9c2956d2 authored by Zhao Chen's avatar Zhao Chen Committed by David S. Miller

net-next: hinic: fix a problem in free_tx_poll()

This patch fixes the problem below. The problem can be reproduced by the
following steps:
1) Connecting all HiNIC interfaces
2) On server side
    # sudo ifconfig eth0 192.168.100.1 up #Using MLX CX4 card
    # iperf -s
3) On client side
    # sudo ifconfig eth0 192.168.100.2 up #Using our HiNIC card
    # iperf -c 192.168.101.1 -P 10 -t 100000

after hours of testing, we will see errors:

    hinic 0000:05:00.0: No MGMT msg handler, mod = 0
    hinic 0000:05:00.0: No MGMT msg handler, mod = 0
    hinic 0000:05:00.0: No MGMT msg handler, mod = 0
    hinic 0000:05:00.0: No MGMT msg handler, mod = 0

The errors are caused by the following problem.
1) The hinic_get_wqe() checks the "wq->delta" to allocate new WQEs:

	if (atomic_sub_return(num_wqebbs, &wq->delta) <= 0) {
		atomic_add(num_wqebbs, &wq->delta);
		return ERR_PTR(-EBUSY);
	}

If the WQE occupies multiple pages, the shadow WQE will be used. Then the
hinic_xmit_frame() fills the WQE.

2) While in parallel with 1), the free_tx_poll() checks the "wq->delta"
to free old WQEs:

	if ((atomic_read(&wq->delta) + num_wqebbs) > wq->q_depth)
		return ERR_PTR(-EBUSY);

There is a probability that the shadow WQE which hinic_xmit_frame() is
using will be damaged by copy_wqe_to_shadow():

	if (curr_pg != end_pg) {
		void *shadow_addr = &wq->shadow_wqe[curr_pg * wq->max_wqe_size];

		copy_wqe_to_shadow(wq, shadow_addr, num_wqebbs, *cons_idx);
		return shadow_addr;
	}

This can cause WQE data error and you will see the above error messages.
This patch fixes the problem.
Signed-off-by: default avatarZhao Chen <zhaochen6@huawei.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent 5a0c6cee
......@@ -635,15 +635,16 @@ void hinic_sq_write_wqe(struct hinic_sq *sq, u16 prod_idx,
}
/**
* hinic_sq_read_wqe - read wqe ptr in the current ci and update the ci
* hinic_sq_read_wqebb - read wqe ptr in the current ci and update the ci, the
* wqe only have one wqebb
* @sq: send queue
* @skb: return skb that was saved
* @wqe_size: the size of the wqe
* @wqe_size: the wqe size ptr
* @cons_idx: consumer index of the wqe
*
* Return wqe in ci position
**/
struct hinic_sq_wqe *hinic_sq_read_wqe(struct hinic_sq *sq,
struct hinic_sq_wqe *hinic_sq_read_wqebb(struct hinic_sq *sq,
struct sk_buff **skb,
unsigned int *wqe_size, u16 *cons_idx)
{
......@@ -658,6 +659,8 @@ struct hinic_sq_wqe *hinic_sq_read_wqe(struct hinic_sq *sq,
if (IS_ERR(hw_wqe))
return NULL;
*skb = sq->saved_skb[*cons_idx];
sq_wqe = &hw_wqe->sq_wqe;
ctrl = &sq_wqe->ctrl;
ctrl_info = be32_to_cpu(ctrl->ctrl_info);
......@@ -665,11 +668,28 @@ struct hinic_sq_wqe *hinic_sq_read_wqe(struct hinic_sq *sq,
*wqe_size = sizeof(*ctrl) + sizeof(sq_wqe->task);
*wqe_size += SECT_SIZE_FROM_8BYTES(buf_sect_len);
*wqe_size = ALIGN(*wqe_size, sq->wq->wqebb_size);
*skb = sq->saved_skb[*cons_idx];
return &hw_wqe->sq_wqe;
}
/* using the real wqe size to read wqe again */
hw_wqe = hinic_read_wqe(sq->wq, *wqe_size, cons_idx);
/**
* hinic_sq_read_wqe - read wqe ptr in the current ci and update the ci
* @sq: send queue
* @skb: return skb that was saved
* @wqe_size: the size of the wqe
* @cons_idx: consumer index of the wqe
*
* Return wqe in ci position
**/
struct hinic_sq_wqe *hinic_sq_read_wqe(struct hinic_sq *sq,
struct sk_buff **skb,
unsigned int wqe_size, u16 *cons_idx)
{
struct hinic_hw_wqe *hw_wqe;
hw_wqe = hinic_read_wqe(sq->wq, wqe_size, cons_idx);
*skb = sq->saved_skb[*cons_idx];
return &hw_wqe->sq_wqe;
}
......
......@@ -164,6 +164,10 @@ void hinic_sq_write_wqe(struct hinic_sq *sq, u16 prod_idx,
unsigned int wqe_size);
struct hinic_sq_wqe *hinic_sq_read_wqe(struct hinic_sq *sq,
struct sk_buff **skb,
unsigned int wqe_size, u16 *cons_idx);
struct hinic_sq_wqe *hinic_sq_read_wqebb(struct hinic_sq *sq,
struct sk_buff **skb,
unsigned int *wqe_size, u16 *cons_idx);
......
......@@ -283,7 +283,11 @@ static void free_all_tx_skbs(struct hinic_txq *txq)
int nr_sges;
u16 ci;
while ((sq_wqe = hinic_sq_read_wqe(sq, &skb, &wqe_size, &ci))) {
while ((sq_wqe = hinic_sq_read_wqebb(sq, &skb, &wqe_size, &ci))) {
sq_wqe = hinic_sq_read_wqe(sq, &skb, wqe_size, &ci);
if (!sq_wqe)
break;
nr_sges = skb_shinfo(skb)->nr_frags + 1;
hinic_sq_get_sges(sq_wqe, txq->free_sges, nr_sges);
......@@ -319,11 +323,21 @@ static int free_tx_poll(struct napi_struct *napi, int budget)
do {
hw_ci = HW_CONS_IDX(sq) & wq->mask;
sq_wqe = hinic_sq_read_wqe(sq, &skb, &wqe_size, &sw_ci);
/* Reading a WQEBB to get real WQE size and consumer index. */
sq_wqe = hinic_sq_read_wqebb(sq, &skb, &wqe_size, &sw_ci);
if ((!sq_wqe) ||
(((hw_ci - sw_ci) & wq->mask) * wq->wqebb_size < wqe_size))
break;
/* If this WQE have multiple WQEBBs, we will read again to get
* full size WQE.
*/
if (wqe_size > wq->wqebb_size) {
sq_wqe = hinic_sq_read_wqe(sq, &skb, wqe_size, &sw_ci);
if (unlikely(!sq_wqe))
break;
}
tx_bytes += skb->len;
pkts++;
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment