Skip to contentSkip to conversation

hinic: BugFixes

PATCH In review Luo bin 3 participants 11 messages Mar 16, 2020 – Mar 19, 2020

Subsystems (3)
lore
Share & tools
lore

Luo bin

author
#0 lore raw

cover hinic: BugFixes

Fix a number of bugs which have been present since the first commit.

The bugs fixed in these patchs are hardly exposed unless given very specific conditions.

Luo bin (6):
hinic: fix process of long length skb without frags
hinic: fix a bug of waitting for IO stopped
hinic: fix the bug of clearing event queue
hinic: fix out-of-order excution in arm cpu
hinic: fix wrong para of wait_for_completion_timeout
hinic: fix wrong value of MIN_SKB_LEN

 .../net/ethernet/huawei/hinic/hinic_ethtool.c |  1 +
 .../net/ethernet/huawei/hinic/hinic_hw_cmdq.c |  5 +-
 .../net/ethernet/huawei/hinic/hinic_hw_dev.c  | 51 +------------------
 .../net/ethernet/huawei/hinic/hinic_hw_eqs.c  | 27 +++++++---
 .../net/ethernet/huawei/hinic/hinic_hw_mgmt.c |  5 +-
 .../net/ethernet/huawei/hinic/hinic_main.c    |  1 +
 drivers/net/ethernet/huawei/hinic/hinic_rx.c  |  3 ++
 drivers/net/ethernet/huawei/hinic/hinic_tx.c  | 15 ++++--
 drivers/net/ethernet/huawei/hinic/hinic_tx.h  |  2 +-
 9 files changed, 46 insertions(+), 64 deletions(-)
Signature · 1 line
2.17.1
author
#1 lore raw

PATCH 1/6 hinic: fix process of long length skb without frags

some tool such as pktgen can build an illegal skb with long length but no fragments, which is unsupported for hw, so drop it

---
 drivers/net/ethernet/huawei/hinic/hinic_ethtool.c |  1 +
 drivers/net/ethernet/huawei/hinic/hinic_main.c    |  1 +
 drivers/net/ethernet/huawei/hinic/hinic_tx.c      | 13 +++++++++----
 drivers/net/ethernet/huawei/hinic/hinic_tx.h      |  2 +-
 4 files changed, 12 insertions(+), 5 deletions(-)
drivers/net/ethernet/huawei/hinic/hinic_ethtool.c +1 -0
index 966aea949c0b..dac157bc06a7 100644
@@ -582,6 +582,7 @@ static struct hinic_stats hinic_tx_queue_stats[] = {
 	HINIC_TXQ_STAT(tx_wake),
 	HINIC_TXQ_STAT(tx_dropped),
 	HINIC_TXQ_STAT(big_frags_pkts),
+	HINIC_TXQ_STAT(frag_len_overflow),
 };
 
 #define HINIC_RXQ_STAT(_stat_item) { \
drivers/net/ethernet/huawei/hinic/hinic_main.c +1 -0
index 13560975c103..a9bee70bd6c7 100644
@@ -107,6 +107,7 @@ static void update_tx_stats(struct hinic_dev *nic_dev, struct hinic_txq *txq)
 	nic_tx_stats->tx_wake += tx_stats.tx_wake;
 	nic_tx_stats->tx_dropped += tx_stats.tx_dropped;
 	nic_tx_stats->big_frags_pkts += tx_stats.big_frags_pkts;
+	nic_tx_stats->frag_len_overflow += tx_stats.frag_len_overflow;
 	u64_stats_update_end(&nic_tx_stats->syncp);
 
 	hinic_txq_clean_stats(txq);
drivers/net/ethernet/huawei/hinic/hinic_tx.c +9 -4
index 0e13d1c7e474..3c6762086fff 100644
@@ -45,9 +45,10 @@
 
 #define HW_CONS_IDX(sq)                 be16_to_cpu(*(u16 *)((sq)->hw_ci_addr))
 
-#define MIN_SKB_LEN                     17
+#define MIN_SKB_LEN			17
+#define HINIC_GSO_MAX_SIZE		65536
 
-#define	MAX_PAYLOAD_OFFSET	        221
+#define	MAX_PAYLOAD_OFFSET		221
 #define TRANSPORT_OFFSET(l4_hdr, skb)	((u32)((l4_hdr) - (skb)->data))
 
 union hinic_l3 {
@@ -84,6 +85,7 @@ void hinic_txq_clean_stats(struct hinic_txq *txq)
 	txq_stats->tx_wake = 0;
 	txq_stats->tx_dropped = 0;
 	txq_stats->big_frags_pkts = 0;
+	txq_stats->frag_len_overflow = 0;
 	u64_stats_update_end(&txq_stats->syncp);
 }
 
@@ -106,6 +108,7 @@ void hinic_txq_get_stats(struct hinic_txq *txq, struct hinic_txq_stats *stats)
 		stats->tx_wake = txq_stats->tx_wake;
 		stats->tx_dropped = txq_stats->tx_dropped;
 		stats->big_frags_pkts = txq_stats->big_frags_pkts;
+		stats->frag_len_overflow = txq_stats->frag_len_overflow;
 	} while (u64_stats_fetch_retry(&txq_stats->syncp, start));
 	u64_stats_update_end(&stats->syncp);
 }
@@ -440,7 +443,6 @@ static int hinic_tx_offload(struct sk_buff *skb, struct hinic_sq_task *task,
 			     vlan_tag >> VLAN_PRIO_SHIFT);
 		offload |= TX_OFFLOAD_VLAN;
 	}
-
 	if (offload)
 		hinic_task_set_l2hdr(task, skb_network_offset(skb));
 
@@ -488,11 +490,14 @@ netdev_tx_t hinic_xmit_frame(struct sk_buff *skb, struct net_device *netdev)
 		txq->txq_stats.big_frags_pkts++;
 		u64_stats_update_end(&txq->txq_stats.syncp);
 	}
-
 	if (nr_sges > txq->max_sges) {
 		netdev_err(netdev, "Too many Tx sges\n");
 		goto skb_error;
 	}
+	if (unlikely(skb->len > HINIC_GSO_MAX_SIZE && nr_sges == 1)) {
+		txq->txq_stats.frag_len_overflow++;
+		goto skb_error;
+	}
 
 	err = tx_map_skb(nic_dev, skb, txq->sges);
 	if (err)
drivers/net/ethernet/huawei/hinic/hinic_tx.h +1 -1
index f158b7db7fb8..ac65b4301c09 100644
@@ -22,7 +22,7 @@ struct hinic_txq_stats {
 	u64     tx_wake;
 	u64     tx_dropped;
 	u64	big_frags_pkts;
-
+	u64     frag_len_overflow;
 	struct u64_stats_sync   syncp;
 };
 
Signature · 1 line
2.17.1
maintainer
#2 lore raw

Re: [PATCH net 1/6] hinic: fix process of long length skb without frags

drivers/net/ethernet/huawei/hinic/hinic_tx.c

Quoted 8 lines · On Mon, 16 Mar 2020 00:56:25 +0000 Luo bin wrote:
drivers/net/ethernet/huawei/hinic/hinic_tx.c +6 -1
-#define MIN_SKB_LEN                     17
+#define MIN_SKB_LEN			17
+#define HINIC_GSO_MAX_SIZE		65536
 
+	if (unlikely(skb->len > HINIC_GSO_MAX_SIZE && nr_sges == 1)) {
+		txq->txq_stats.frag_len_overflow++;
+		goto skb_error;
+	}

I don't think drivers should have to check this condition.

We have netdev->gso_max_size which should be initialized to

include/linux/netdevice.h:#define GSO_MAX_SIZE 65536

in

net/core/dev.c: dev->gso_max_size = GSO_MAX_SIZE;

Please send a patch to pktgen to uphold the normal stack guarantees.

maintainer
#3 lore raw

Re: [PATCH net 1/6] hinic: fix process of long length skb without frags

From: Jakub Kicinski <[email protected]>
Date: Mon, 16 Mar 2020 14:44:08 -0700

Quoted 14 lines · On Mon, 16 Mar 2020 00:56:25 +0000 Luo bin wrote:

drivers/net/ethernet/huawei/hinic/hinic_tx.c

Quoted 8 lines
drivers/net/ethernet/huawei/hinic/hinic_tx.c +6 -1
-#define MIN_SKB_LEN                     17
+#define MIN_SKB_LEN			17
+#define HINIC_GSO_MAX_SIZE		65536
 
+	if (unlikely(skb->len > HINIC_GSO_MAX_SIZE && nr_sges == 1)) {
+		txq->txq_stats.frag_len_overflow++;
+		goto skb_error;
+	}

I don't think drivers should have to check this condition.

We have netdev->gso_max_size which should be initialized to

Quoted 7 lines

include/linux/netdevice.h:#define GSO_MAX_SIZE 65536

in

net/core/dev.c: dev->gso_max_size = GSO_MAX_SIZE;

Please send a patch to pktgen to uphold the normal stack guarantees.

Agreed, the driver should not have to validate this.

author
#4 lore raw

Re: [PATCH net 1/6] hinic: fix process of long length skb without frags

Ok,I will undo this patch.

Quoted 24 lines · David Miller

From: Jakub Kicinski <[email protected]>
Date: Mon, 16 Mar 2020 14:44:08 -0700

Quoted 12 lines · On Mon, 16 Mar 2020 00:56:25 +0000 Luo bin wrote:

drivers/net/ethernet/huawei/hinic/hinic_tx.c netdev_tx_t hinic_xmit_frame(struct sk_buff *skb, struct net_device *netdev) · line 500

Quoted 7 lines
drivers/net/ethernet/huawei/hinic/hinic_tx.c +6 -1
-#define MIN_SKB_LEN                     17
+#define MIN_SKB_LEN			17
+#define HINIC_GSO_MAX_SIZE		65536
+	if (unlikely(skb->len > HINIC_GSO_MAX_SIZE && nr_sges == 1)) {
+		txq->txq_stats.frag_len_overflow++;
+		goto skb_error;
+	}

I don't think drivers should have to check this condition.

We have netdev->gso_max_size which should be initialized to

Quoted 7 lines

include/linux/netdevice.h:#define GSO_MAX_SIZE 65536

in

net/core/dev.c: dev->gso_max_size = GSO_MAX_SIZE;

Please send a patch to pktgen to uphold the normal stack guarantees.

Agreed, the driver should not have to validate this. .

author
#5 lore raw

PATCH 2/6 hinic: fix a bug of waitting for IO stopped

In reply to Luo bin

it's unreliable for fw to check whether IO is stopped, so driver wait for enough time to ensure IO process is done in hw before freeing resources

---
 .../net/ethernet/huawei/hinic/hinic_hw_dev.c  | 51 +------------------
 1 file changed, 2 insertions(+), 49 deletions(-)
drivers/net/ethernet/huawei/hinic/hinic_hw_dev.c +2 -49
index 79b3d53f2fbf..c7c75b772a86 100644
@@ -360,50 +360,6 @@ static int wait_for_db_state(struct hinic_hwdev *hwdev)
 	return -EFAULT;
 }
 
-static int wait_for_io_stopped(struct hinic_hwdev *hwdev)
-{
-	struct hinic_cmd_io_status cmd_io_status;
-	struct hinic_hwif *hwif = hwdev->hwif;
-	struct pci_dev *pdev = hwif->pdev;
-	struct hinic_pfhwdev *pfhwdev;
-	unsigned long end;
-	u16 out_size;
-	int err;
-
-	if (!HINIC_IS_PF(hwif) && !HINIC_IS_PPF(hwif)) {
-		dev_err(&pdev->dev, "Unsupported PCI Function type\n");
-		return -EINVAL;
-	}
-
-	pfhwdev = container_of(hwdev, struct hinic_pfhwdev, hwdev);
-
-	cmd_io_status.func_idx = HINIC_HWIF_FUNC_IDX(hwif);
-
-	end = jiffies + msecs_to_jiffies(IO_STATUS_TIMEOUT);
-	do {
-		err = hinic_msg_to_mgmt(&pfhwdev->pf_to_mgmt, HINIC_MOD_COMM,
-					HINIC_COMM_CMD_IO_STATUS_GET,
-					&cmd_io_status, sizeof(cmd_io_status),
-					&cmd_io_status, &out_size,
-					HINIC_MGMT_MSG_SYNC);
-		if ((err) || (out_size != sizeof(cmd_io_status))) {
-			dev_err(&pdev->dev, "Failed to get IO status, ret = %d\n",
-				err);
-			return err;
-		}
-
-		if (cmd_io_status.status == IO_STOPPED) {
-			dev_info(&pdev->dev, "IO stopped\n");
-			return 0;
-		}
-
-		msleep(20);
-	} while (time_before(jiffies, end));
-
-	dev_err(&pdev->dev, "Wait for IO stopped - Timeout\n");
-	return -ETIMEDOUT;
-}
-
 /**
  * clear_io_resource - set the IO resources as not active in the NIC
  * @hwdev: the NIC HW device
@@ -423,11 +379,8 @@ static int clear_io_resources(struct hinic_hwdev *hwdev)
 		return -EINVAL;
 	}
 
-	err = wait_for_io_stopped(hwdev);
-	if (err) {
-		dev_err(&pdev->dev, "IO has not stopped yet\n");
-		return err;
-	}
+	/* sleep 100ms to wait for firmware stopping I/O */
+	msleep(100);
 
 	cmd_clear_io_res.func_idx = HINIC_HWIF_FUNC_IDX(hwif);
 
Signature · 1 line
2.17.1
author
#6 lore raw

PATCH 3/6 hinic: fix the bug of clearing event queue

In reply to Luo bin

should disable and synchronize eq irq before freeing it, must clear event queue depth in hw before freeing relevant memory to avoid illegal memory access and update consumer idx to avoid invalid interrupt

---
 .../net/ethernet/huawei/hinic/hinic_hw_eqs.c  | 25 +++++++++++++------
 1 file changed, 18 insertions(+), 7 deletions(-)
drivers/net/ethernet/huawei/hinic/hinic_hw_eqs.c +18 -7
index 79243b626ddb..13f3abac54ea 100644
@@ -188,7 +188,7 @@ static u8 eq_cons_idx_checksum_set(u32 val)
  * eq_update_ci - update the HW cons idx of event queue
  * @eq: the event queue to update the cons idx for
  **/
-static void eq_update_ci(struct hinic_eq *eq)
+static void eq_update_ci(struct hinic_eq *eq, u32 arm_state)
 {
 	u32 val, addr = EQ_CONS_IDX_REG_ADDR(eq);
 
@@ -202,7 +202,7 @@ static void eq_update_ci(struct hinic_eq *eq)
 
 	val |= HINIC_EQ_CI_SET(eq->cons_idx, IDX)    |
 	       HINIC_EQ_CI_SET(eq->wrapped, WRAPPED) |
-	       HINIC_EQ_CI_SET(EQ_ARMED, INT_ARMED);
+	       HINIC_EQ_CI_SET(arm_state, INT_ARMED);
 
 	val |= HINIC_EQ_CI_SET(eq_cons_idx_checksum_set(val), XOR_CHKSUM);
 
@@ -347,7 +347,7 @@ static void eq_irq_handler(void *data)
 	else if (eq->type == HINIC_CEQ)
 		ceq_irq_handler(eq);
 
-	eq_update_ci(eq);
+	eq_update_ci(eq, EQ_ARMED);
 }
 
 /**
@@ -702,7 +702,7 @@ static int init_eq(struct hinic_eq *eq, struct hinic_hwif *hwif,
 	}
 
 	set_eq_ctrls(eq);
-	eq_update_ci(eq);
+	eq_update_ci(eq, EQ_ARMED);
 
 	err = alloc_eq_pages(eq);
 	if (err) {
@@ -752,18 +752,29 @@ static int init_eq(struct hinic_eq *eq, struct hinic_hwif *hwif,
  **/
 static void remove_eq(struct hinic_eq *eq)
 {
-	struct msix_entry *entry = &eq->msix_entry;
-
-	free_irq(entry->vector, eq);
+	hinic_set_msix_state(eq->hwif, eq->msix_entry.entry,
+			     HINIC_MSIX_DISABLE);
+	synchronize_irq(eq->msix_entry.vector);
+	free_irq(eq->msix_entry.vector, eq);
 
 	if (eq->type == HINIC_AEQ) {
 		struct hinic_eq_work *aeq_work = &eq->aeq_work;
 
 		cancel_work_sync(&aeq_work->work);
+		/* clear aeq_len to avoid hw access host memory */
+		hinic_hwif_write_reg(eq->hwif,
+				     HINIC_CSR_AEQ_CTRL_1_ADDR(eq->q_id), 0);
 	} else if (eq->type == HINIC_CEQ) {
 		tasklet_kill(&eq->ceq_tasklet);
+		/* clear ceq_len to avoid hw access host memory */
+		hinic_hwif_write_reg(eq->hwif,
+				     HINIC_CSR_CEQ_CTRL_1_ADDR(eq->q_id), 0);
 	}
 
+	/* update cons_idx to avoid invalid interrupt */
+	eq->cons_idx = hinic_hwif_read_reg(eq->hwif, EQ_PROD_IDX_REG_ADDR(eq));
+	eq_update_ci(eq, EQ_NOT_ARMED);
+
 	free_eq_pages(eq);
 }
 
Signature · 1 line
2.17.1
maintainer
#7 lore raw

Re: [PATCH net 3/6] hinic: fix the bug of clearing event queue

drivers/net/ethernet/huawei/hinic/hinic_hw_eqs.c static int init_eq(struct hinic_eq *eq, struct hinic_hwif *hwif, · line 758

Quoted 2 lines · On Mon, 16 Mar 2020 00:56:27 +0000 Luo bin wrote:
drivers/net/ethernet/huawei/hinic/hinic_hw_eqs.c +2 -0
+	synchronize_irq(eq->msix_entry.vector);
+	free_irq(eq->msix_entry.vector, eq);

Doesn't free_irq() imply synchronize_irq()?

author
#8 lore raw

PATCH 4/6 hinic: fix out-of-order excution in arm cpu

In reply to Luo bin

add read barrier in driver code to keep from reading other fileds in dma memory which is writable for hw until we have verified the memory is valid for driver

---
 drivers/net/ethernet/huawei/hinic/hinic_hw_cmdq.c | 2 ++
 drivers/net/ethernet/huawei/hinic/hinic_hw_eqs.c  | 2 ++
 drivers/net/ethernet/huawei/hinic/hinic_rx.c      | 3 +++
 drivers/net/ethernet/huawei/hinic/hinic_tx.c      | 2 ++
 4 files changed, 9 insertions(+)
drivers/net/ethernet/huawei/hinic/hinic_hw_cmdq.c +2 -0
index eb53c15b13f3..33f93cc25193 100644
@@ -623,6 +623,8 @@ static int cmdq_cmd_ceq_handler(struct hinic_cmdq *cmdq, u16 ci,
 	if (!CMDQ_WQE_COMPLETED(be32_to_cpu(ctrl->ctrl_info)))
 		return -EBUSY;
 
+	dma_rmb();
+
 	errcode = CMDQ_WQE_ERRCODE_GET(be32_to_cpu(status->status_info), VAL);
 
 	cmdq_sync_cmd_handler(cmdq, ci, errcode);
drivers/net/ethernet/huawei/hinic/hinic_hw_eqs.c +2 -0
index 13f3abac54ea..fb662ab9b85d 100644
@@ -235,6 +235,8 @@ static void aeq_irq_handler(struct hinic_eq *eq)
 		if (HINIC_EQ_ELEM_DESC_GET(aeqe_desc, WRAPPED) == eq->wrapped)
 			break;
 
+		dma_rmb();
+
 		event = HINIC_EQ_ELEM_DESC_GET(aeqe_desc, TYPE);
 		if (event >= HINIC_MAX_AEQ_EVENTS) {
 			dev_err(&pdev->dev, "Unknown AEQ Event %d\n", event);
drivers/net/ethernet/huawei/hinic/hinic_rx.c +3 -0
index 2695ad69fca6..815649e37cb1 100644
@@ -350,6 +350,9 @@ static int rxq_recv(struct hinic_rxq *rxq, int budget)
 		if (!rq_wqe)
 			break;
 
+		/* make sure we read rx_done before packet length */
+		dma_rmb();
+
 		cqe = rq->cqe[ci];
 		status =  be32_to_cpu(cqe->status);
 		hinic_rq_get_sge(rxq->rq, rq_wqe, ci, &sge);
drivers/net/ethernet/huawei/hinic/hinic_tx.c +2 -0
index 3c6762086fff..cabcc9019ee4 100644
@@ -627,6 +627,8 @@ static int free_tx_poll(struct napi_struct *napi, int budget)
 	do {
 		hw_ci = HW_CONS_IDX(sq) & wq->mask;
 
+		dma_rmb();
+
 		/* 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) ||
Signature · 1 line
2.17.1
author
#9 lore raw

PATCH 5/6 hinic: fix wrong para of wait_for_completion_timeout

In reply to Luo bin

the second input parameter of wait_for_completion_timeout should be jiffies instead of millisecond

---
 drivers/net/ethernet/huawei/hinic/hinic_hw_cmdq.c | 3 ++-
 drivers/net/ethernet/huawei/hinic/hinic_hw_mgmt.c | 5 +++--
 2 files changed, 5 insertions(+), 3 deletions(-)
drivers/net/ethernet/huawei/hinic/hinic_hw_cmdq.c +2 -1
index 33f93cc25193..5f2d57d1b2d3 100644
@@ -389,7 +389,8 @@ static int cmdq_sync_cmd_direct_resp(struct hinic_cmdq *cmdq,
 
 	spin_unlock_bh(&cmdq->cmdq_lock);
 
-	if (!wait_for_completion_timeout(&done, CMDQ_TIMEOUT)) {
+	if (!wait_for_completion_timeout(&done,
+					 msecs_to_jiffies(CMDQ_TIMEOUT))) {
 		spin_lock_bh(&cmdq->cmdq_lock);
 
 		if (cmdq->errcode[curr_prod_idx] == &errcode)
drivers/net/ethernet/huawei/hinic/hinic_hw_mgmt.c +3 -2
index c1a6be6bf6a8..8995e32dd1c0 100644
@@ -43,7 +43,7 @@
 
 #define MSG_NOT_RESP                    0xFFFF
 
-#define MGMT_MSG_TIMEOUT                1000
+#define MGMT_MSG_TIMEOUT                5000
 
 #define mgmt_to_pfhwdev(pf_mgmt)        \
 		container_of(pf_mgmt, struct hinic_pfhwdev, pf_to_mgmt)
@@ -267,7 +267,8 @@ static int msg_to_mgmt_sync(struct hinic_pf_to_mgmt *pf_to_mgmt,
 		goto unlock_sync_msg;
 	}
 
-	if (!wait_for_completion_timeout(recv_done, MGMT_MSG_TIMEOUT)) {
+	if (!wait_for_completion_timeout(recv_done,
+					 msecs_to_jiffies(MGMT_MSG_TIMEOUT))) {
 		dev_err(&pdev->dev, "MGMT timeout, MSG id = %d\n", msg_id);
 		err = -ETIMEDOUT;
 		goto unlock_sync_msg;
Signature · 1 line
2.17.1
author
#10 lore raw

PATCH 6/6 hinic: fix wrong value of MIN_SKB_LEN

In reply to Luo bin

the minimum value of skb len that hw supports is 32 rather than 17

---
 drivers/net/ethernet/huawei/hinic/hinic_tx.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/net/ethernet/huawei/hinic/hinic_tx.c +1 -1
index cabcc9019ee4..8993f5b07059 100644
@@ -45,7 +45,7 @@
 
 #define HW_CONS_IDX(sq)                 be16_to_cpu(*(u16 *)((sq)->hw_ci_addr))
 
-#define MIN_SKB_LEN			17
+#define MIN_SKB_LEN			32
 #define HINIC_GSO_MAX_SIZE		65536
 
 #define	MAX_PAYLOAD_OFFSET		221
Signature · 1 line
2.17.1