[PATCH net 0/2] bnxt_en: Add workaround to detect bad opaque in rx completion.

[PATCH net 0/2] bnxt_en: Add workaround to detect bad opaque in rx completion.

[Michael Chan]

2-part workaround for this hardware bug.

Michael Chan (2):
  bnxt_en: Add workaround to detect bad opaque in rx completion (part 1)
  bnxt_en: Add workaround to detect bad opaque in rx completion (part 2)

 drivers/net/ethernet/broadcom/bnxt/bnxt.c | 63 +++++++++++++++++++++++++++++++
 drivers/net/ethernet/broadcom/bnxt/bnxt.h |  2 +
 2 files changed, 65 insertions(+)

-- 
1.8.3.1

[PATCH net 1/2] bnxt_en: Add workaround to detect bad opaque in rx completion (part 1)

[Michael Chan]

There is a rare hardware bug that can cause a bad opaque value in the RX
or TPA completion.  When this happens, the hardware may have used the
same buffer twice for 2 rx packets.  In addition, the driver will also
crash later using the bad opaque as the index into the ring.

The rx opaque value is predictable and is always monotonically increasing.
The workaround is to keep track of the expected next opaque value and
compare it with the one returned by hardware during RX and TPA start
completions.  If they miscompare, we will not process any more RX and
TPA completions and exit NAPI.  We will then schedule a workqueue to
reset the function.

This patch adds the logic to keep track of the next rx consumer index.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
---
 drivers/net/ethernet/broadcom/bnxt/bnxt.c | 3 +++
 drivers/net/ethernet/broadcom/bnxt/bnxt.h | 1 +
 2 files changed, 4 insertions(+)

diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
index 9d4e8e1..58999cd 100644
--- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
+++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
@@ -867,6 +867,7 @@ static void bnxt_tpa_start(struct bnxt *bp, struct bnxt_rx_ring_info *rxr,
 
 	rxr->rx_prod = NEXT_RX(prod);
 	cons = NEXT_RX(cons);
+	rxr->rx_next_cons = NEXT_RX(cons);
 	cons_rx_buf = &rxr->rx_buf_ring[cons];
 
 	bnxt_reuse_rx_data(rxr, cons, cons_rx_buf->data);
@@ -1245,6 +1246,7 @@ static int bnxt_rx_pkt(struct bnxt *bp, struct bnxt_napi *bnapi, u32 *raw_cons,
 
 next_rx:
 	rxr->rx_prod = NEXT_RX(prod);
+	rxr->rx_next_cons = NEXT_RX(cons);
 
 next_rx_no_prod:
 	*raw_cons = tmp_raw_cons;
@@ -2486,6 +2488,7 @@ static void bnxt_clear_ring_indices(struct bnxt *bp)
 			rxr->rx_prod = 0;
 			rxr->rx_agg_prod = 0;
 			rxr->rx_sw_agg_prod = 0;
+			rxr->rx_next_cons = 0;
 		}
 	}
 }
diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.h b/drivers/net/ethernet/broadcom/bnxt/bnxt.h
index 8b823ff..52f2d74 100644
--- a/drivers/net/ethernet/broadcom/bnxt/bnxt.h
+++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.h
@@ -584,6 +584,7 @@ struct bnxt_rx_ring_info {
 	u16			rx_prod;
 	u16			rx_agg_prod;
 	u16			rx_sw_agg_prod;
+	u16			rx_next_cons;
 	void __iomem		*rx_doorbell;
 	void __iomem		*rx_agg_doorbell;
 
-- 
1.8.3.1

[PATCH net 2/2] bnxt_en: Add workaround to detect bad opaque in rx completion (part 2)

[Michael Chan]

Add detection and recovery code when the hardware returned opaque value
does not match the expected consumer index.  Once the issue is detected,
we skip the processing of all RX and LRO/GRO packets.  These completion
entries are discarded without sending the SKB to the stack and without
producing new buffers.  The function will be reset from a workqueue.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
---
 drivers/net/ethernet/broadcom/bnxt/bnxt.c | 60 +++++++++++++++++++++++++++++++
 drivers/net/ethernet/broadcom/bnxt/bnxt.h |  1 +
 2 files changed, 61 insertions(+)

diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
index 58999cd..c39a7f5 100644
--- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
+++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
@@ -813,6 +813,46 @@ static inline struct sk_buff *bnxt_copy_skb(struct bnxt_napi *bnapi, u8 *data,
 	return skb;
 }
 
+static int bnxt_discard_rx(struct bnxt *bp, struct bnxt_napi *bnapi,
+			   u32 *raw_cons, void *cmp)
+{
+	struct bnxt_cp_ring_info *cpr = &bnapi->cp_ring;
+	struct rx_cmp *rxcmp = cmp;
+	u32 tmp_raw_cons = *raw_cons;
+	u8 cmp_type, agg_bufs = 0;
+
+	cmp_type = RX_CMP_TYPE(rxcmp);
+
+	if (cmp_type == CMP_TYPE_RX_L2_CMP) {
+		agg_bufs = (le32_to_cpu(rxcmp->rx_cmp_misc_v1) &
+			    RX_CMP_AGG_BUFS) >>
+			   RX_CMP_AGG_BUFS_SHIFT;
+	} else if (cmp_type == CMP_TYPE_RX_L2_TPA_END_CMP) {
+		struct rx_tpa_end_cmp *tpa_end = cmp;
+
+		agg_bufs = (le32_to_cpu(tpa_end->rx_tpa_end_cmp_misc_v1) &
+			    RX_TPA_END_CMP_AGG_BUFS) >>
+			   RX_TPA_END_CMP_AGG_BUFS_SHIFT;
+	}
+
+	if (agg_bufs) {
+		if (!bnxt_agg_bufs_valid(bp, cpr, agg_bufs, &tmp_raw_cons))
+			return -EBUSY;
+	}
+	*raw_cons = tmp_raw_cons;
+	return 0;
+}
+
+static void bnxt_sched_reset(struct bnxt *bp, struct bnxt_rx_ring_info *rxr)
+{
+	if (!rxr->bnapi->in_reset) {
+		rxr->bnapi->in_reset = true;
+		set_bit(BNXT_RESET_TASK_SP_EVENT, &bp->sp_event);
+		schedule_work(&bp->sp_task);
+	}
+	rxr->rx_next_cons = 0xffff;
+}
+
 static void bnxt_tpa_start(struct bnxt *bp, struct bnxt_rx_ring_info *rxr,
 			   struct rx_tpa_start_cmp *tpa_start,
 			   struct rx_tpa_start_cmp_ext *tpa_start1)
@@ -830,6 +870,11 @@ static void bnxt_tpa_start(struct bnxt *bp, struct bnxt_rx_ring_info *rxr,
 	prod_rx_buf = &rxr->rx_buf_ring[prod];
 	tpa_info = &rxr->rx_tpa[agg_id];
 
+	if (unlikely(cons != rxr->rx_next_cons)) {
+		bnxt_sched_reset(bp, rxr);
+		return;
+	}
+
 	prod_rx_buf->data = tpa_info->data;
 
 	mapping = tpa_info->mapping;
@@ -981,6 +1026,14 @@ static inline struct sk_buff *bnxt_tpa_end(struct bnxt *bp,
 	dma_addr_t mapping;
 	struct sk_buff *skb;
 
+	if (unlikely(bnapi->in_reset)) {
+		int rc = bnxt_discard_rx(bp, bnapi, raw_cons, tpa_end);
+
+		if (rc < 0)
+			return ERR_PTR(-EBUSY);
+		return NULL;
+	}
+
 	tpa_info = &rxr->rx_tpa[agg_id];
 	data = tpa_info->data;
 	prefetch(data);
@@ -1147,6 +1200,12 @@ static int bnxt_rx_pkt(struct bnxt *bp, struct bnxt_napi *bnapi, u32 *raw_cons,
 	cons = rxcmp->rx_cmp_opaque;
 	rx_buf = &rxr->rx_buf_ring[cons];
 	data = rx_buf->data;
+	if (unlikely(cons != rxr->rx_next_cons)) {
+		int rc1 = bnxt_discard_rx(bp, bnapi, raw_cons, rxcmp);
+
+		bnxt_sched_reset(bp, rxr);
+		return rc1;
+	}
 	prefetch(data);
 
 	agg_bufs = (le32_to_cpu(rxcmp->rx_cmp_misc_v1) & RX_CMP_AGG_BUFS) >>
@@ -4465,6 +4524,7 @@ static void bnxt_enable_napi(struct bnxt *bp)
 	int i;
 
 	for (i = 0; i < bp->cp_nr_rings; i++) {
+		bp->bnapi[i]->in_reset = false;
 		bnxt_enable_poll(bp->bnapi[i]);
 		napi_enable(&bp->bnapi[i]->napi);
 	}
diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.h b/drivers/net/ethernet/broadcom/bnxt/bnxt.h
index 52f2d74..de9d53e 100644
--- a/drivers/net/ethernet/broadcom/bnxt/bnxt.h
+++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.h
@@ -637,6 +637,7 @@ struct bnxt_napi {
 #ifdef CONFIG_NET_RX_BUSY_POLL
 	atomic_t		poll_state;
 #endif
+	bool			in_reset;
 };
 
 #ifdef CONFIG_NET_RX_BUSY_POLL
-- 
1.8.3.1

Re: [PATCH net 0/2] bnxt_en: Add workaround to detect bad opaque in rx completion.

[David Miller]

From: Michael Chan <michael.chan@broadcom.com>
Date: Tue, 10 May 2016 19:17:58 -0400

> 2-part workaround for this hardware bug.

Series applied.