[PATCH] bnx2x: EEH recovery fix

[PATCH] bnx2x: EEH recovery fix

[Eilon Greenstein]

[-- Attachment #1: Type: text/plain, Size: 367 bytes --]

Hi Dave,

The EEH mechanism caused a fatal error when using the bnx2x since the driver attempted accessing the device after it was reset by the EEH. This patch sets a proper unload when EEH error is detected without accessing the device.

Since the current implementation causes a fatal error, I ask that you will consider applying this patch.

Thanks,
Eilon

[-- Attachment #2: 0001-EEH-unload.patch --]
[-- Type: application/octet-stream, Size: 5200 bytes --]

When EEH detects an i/o error it resets the device thus it cannot be accessed.
In this case the driver needs to unload its interface only with OS, kernel and
network stack but not with the device.
After successful recovery, the driver can load normally.

Signed-off-by: Yitchak Gertner <gertner@broadcom.com>
Signed-off-by: Eilon Greenstein <eilong@broadcom.com>
---
 drivers/net/bnx2x_main.c |   95 ++++++++++++++++++++++++++++++++++++++++------
 1 files changed, 83 insertions(+), 12 deletions(-)

diff --git a/drivers/net/bnx2x_main.c b/drivers/net/bnx2x_main.c
index a8eb3c4..64649df 100644
--- a/drivers/net/bnx2x_main.c
+++ b/drivers/net/bnx2x_main.c
@@ -59,8 +59,8 @@
 #include "bnx2x.h"
 #include "bnx2x_init.h"
 
-#define DRV_MODULE_VERSION	"1.45.21"
-#define DRV_MODULE_RELDATE	"2008/09/03"
+#define DRV_MODULE_VERSION	"1.45.22"
+#define DRV_MODULE_RELDATE	"2008/09/07"
 #define BNX2X_BC_VER		0x040200
 
 /* Time in jiffies before concluding the transmitter is hung */
@@ -649,15 +649,16 @@ static void bnx2x_int_disable(struct bnx2x *bp)
 		BNX2X_ERR("BUG! proper val not read from IGU!\n");
 }
 
-static void bnx2x_int_disable_sync(struct bnx2x *bp)
+static void bnx2x_int_disable_sync(struct bnx2x *bp, int disable_hw)
 {
 	int msix = (bp->flags & USING_MSIX_FLAG) ? 1 : 0;
 	int i;
 
 	/* disable interrupt handling */
 	atomic_inc(&bp->intr_sem);
-	/* prevent the HW from sending interrupts */
-	bnx2x_int_disable(bp);
+	if (disable_hw)
+		/* prevent the HW from sending interrupts */
+		bnx2x_int_disable(bp);
 
 	/* make sure all ISRs are done */
 	if (msix) {
@@ -6086,9 +6087,9 @@ static void bnx2x_netif_start(struct bnx2x *bp)
 	}
 }
 
-static void bnx2x_netif_stop(struct bnx2x *bp)
+static void bnx2x_netif_stop(struct bnx2x *bp, int disable_hw)
 {
-	bnx2x_int_disable_sync(bp);
+	bnx2x_int_disable_sync(bp, disable_hw);
 	if (netif_running(bp->dev)) {
 		bnx2x_napi_disable(bp);
 		netif_tx_disable(bp->dev);
@@ -6475,7 +6476,7 @@ load_rings_free:
 	for_each_queue(bp, i)
 		bnx2x_free_rx_sge_range(bp, bp->fp + i, NUM_RX_SGE);
 load_int_disable:
-	bnx2x_int_disable_sync(bp);
+	bnx2x_int_disable_sync(bp, 1);
 	/* Release IRQs */
 	bnx2x_free_irq(bp);
 load_error:
@@ -6650,7 +6651,7 @@ static int bnx2x_nic_unload(struct bnx2x *bp, int unload_mode)
 	bp->rx_mode = BNX2X_RX_MODE_NONE;
 	bnx2x_set_storm_rx_mode(bp);
 
-	bnx2x_netif_stop(bp);
+	bnx2x_netif_stop(bp, 1);
 	if (!netif_running(bp->dev))
 		bnx2x_napi_disable(bp);
 	del_timer_sync(&bp->timer);
@@ -8791,7 +8792,7 @@ static int bnx2x_test_loopback(struct bnx2x *bp, u8 link_up)
 	if (!netif_running(bp->dev))
 		return BNX2X_LOOPBACK_FAILED;
 
-	bnx2x_netif_stop(bp);
+	bnx2x_netif_stop(bp, 1);
 
 	if (bnx2x_run_loopback(bp, BNX2X_MAC_LOOPBACK, link_up)) {
 		DP(NETIF_MSG_PROBE, "MAC loopback failed\n");
@@ -10346,6 +10347,74 @@ static int bnx2x_resume(struct pci_dev *pdev)
 	return rc;
 }
 
+static int bnx2x_eeh_nic_unload(struct bnx2x *bp)
+{
+	int i;
+
+	bp->state = BNX2X_STATE_ERROR;
+
+	bp->rx_mode = BNX2X_RX_MODE_NONE;
+
+	bnx2x_netif_stop(bp, 0);
+
+	del_timer_sync(&bp->timer);
+	bp->stats_state = STATS_STATE_DISABLED;
+	DP(BNX2X_MSG_STATS, "stats_state - DISABLED\n");
+
+	/* Release IRQs */
+	bnx2x_free_irq(bp);
+
+	if (CHIP_IS_E1(bp)) {
+		struct mac_configuration_cmd *config =
+						bnx2x_sp(bp, mcast_config);
+
+		for (i = 0; i < config->hdr.length_6b; i++)
+			CAM_INVALIDATE(config->config_table[i]);
+	}
+
+	/* Free SKBs, SGEs, TPA pool and driver internals */
+	bnx2x_free_skbs(bp);
+	for_each_queue(bp, i)
+		bnx2x_free_rx_sge_range(bp, bp->fp + i, NUM_RX_SGE);
+	bnx2x_free_mem(bp);
+
+	bp->state = BNX2X_STATE_CLOSED;
+
+	netif_carrier_off(bp->dev);
+
+	return 0;
+}
+
+static void bnx2x_eeh_recover(struct bnx2x *bp)
+{
+	u32 val;
+
+	mutex_init(&bp->port.phy_mutex);
+
+	bp->common.shmem_base = REG_RD(bp, MISC_REG_SHARED_MEM_ADDR);
+	bp->link_params.shmem_base = bp->common.shmem_base;
+	BNX2X_DEV_INFO("shmem offset is 0x%x\n", bp->common.shmem_base);
+
+	if (!bp->common.shmem_base ||
+	    (bp->common.shmem_base < 0xA0000) ||
+	    (bp->common.shmem_base >= 0xC0000)) {
+		BNX2X_DEV_INFO("MCP not active\n");
+		bp->flags |= NO_MCP_FLAG;
+		return;
+	}
+
+	val = SHMEM_RD(bp, validity_map[BP_PORT(bp)]);
+	if ((val & (SHR_MEM_VALIDITY_DEV_INFO | SHR_MEM_VALIDITY_MB))
+		!= (SHR_MEM_VALIDITY_DEV_INFO | SHR_MEM_VALIDITY_MB))
+		BNX2X_ERR("BAD MCP validity signature\n");
+
+	if (!BP_NOMCP(bp)) {
+		bp->fw_seq = (SHMEM_RD(bp, func_mb[BP_FUNC(bp)].drv_mb_header)
+			      & DRV_MSG_SEQ_NUMBER_MASK);
+		BNX2X_DEV_INFO("fw_seq 0x%08x\n", bp->fw_seq);
+	}
+}
+
 /**
  * bnx2x_io_error_detected - called when PCI error is detected
  * @pdev: Pointer to PCI device
@@ -10365,7 +10434,7 @@ static pci_ers_result_t bnx2x_io_error_detected(struct pci_dev *pdev,
 	netif_device_detach(dev);
 
 	if (netif_running(dev))
-		bnx2x_nic_unload(bp, UNLOAD_CLOSE);
+		bnx2x_eeh_nic_unload(bp);
 
 	pci_disable_device(pdev);
 
@@ -10420,8 +10489,10 @@ static void bnx2x_io_resume(struct pci_dev *pdev)
 
 	rtnl_lock();
 
+	bnx2x_eeh_recover(bp);
+
 	if (netif_running(dev))
-		bnx2x_nic_load(bp, LOAD_OPEN);
+		bnx2x_nic_load(bp, LOAD_NORMAL);
 
 	netif_device_attach(dev);
 
-- 
1.5.3.2

Re: [PATCH] bnx2x: EEH recovery fix

[David Miller]

From: "Eilon Greenstein" <eilong@broadcom.com>
Date: Mon, 8 Sep 2008 03:22:32 -0700

> The EEH mechanism caused a fatal error when using the bnx2x since
> the driver attempted accessing the device after it was reset by the
> EEH. This patch sets a proper unload when EEH error is detected
> without accessing the device.
>
> Since the current implementation causes a fatal error, I ask that
> you will consider applying this patch.

What user has run into and reported this failure, and is it
a regression compared to 2.6.26?

That is the criteria for my putting this patch in now.

Since EEH triggers are an exceedingly corner case error condition,
this puts this patch even less valid to go in outside of the merge
window.

Re: [PATCH] bnx2x: EEH recovery fix

[Eilon Greenstein]

On Mon, 2008-09-08 at 12:47 -0700, David Miller wrote:
> From: "Eilon Greenstein" <eilong@broadcom.com>
> Date: Mon, 8 Sep 2008 03:22:32 -0700
> 
> > The EEH mechanism caused a fatal error when using the bnx2x since
> > the driver attempted accessing the device after it was reset by the
> > EEH. This patch sets a proper unload when EEH error is detected
> > without accessing the device.
> >
> > Since the current implementation causes a fatal error, I ask that
> > you will consider applying this patch.
> 
> What user has run into and reported this failure, and is it
> a regression compared to 2.6.26?

IBM found it on PPC and it was more obvious with the DMA mapping
violation. It is broken in the same way on 2.6.26 as well.
 
> 
> That is the criteria for my putting this patch in now.
> 
> Since EEH triggers are an exceedingly corner case error condition,
> this puts this patch even less valid to go in outside of the merge
> window.
> 
If you cannot accept it into 2.6.27, I will re-send it into net-next.
Please let me know.

Thanks,
Eilon

Re: [PATCH] bnx2x: EEH recovery fix

[David Miller]

From: "Eilon Greenstein" <eilong@broadcom.com>
Date: Tue, 09 Sep 2008 12:34:49 +0300

> IBM found it on PPC and it was more obvious with the DMA mapping
> violation. It is broken in the same way on 2.6.26 as well.
 ...
> If you cannot accept it into 2.6.27, I will re-send it into net-next.
> Please let me know.

Given what you've said I think net-next is the way to go, please
submit it for there.

Thanks.

Re: [PATCH] bnx2x: EEH recovery fix

[Paul Larson]

David Miller wrote:
> From: "Eilon Greenstein" <eilong@broadcom.com>
> Date: Mon, 8 Sep 2008 03:22:32 -0700
>
>   
>> The EEH mechanism caused a fatal error when using the bnx2x since
>> the driver attempted accessing the device after it was reset by the
>> EEH. This patch sets a proper unload when EEH error is detected
>> without accessing the device.
>>
>> Since the current implementation causes a fatal error, I ask that
>> you will consider applying this patch.
>>     
>
> What user has run into and reported this failure, and is it
> a regression compared to 2.6.26?
>
> That is the criteria for my putting this patch in now.
>
> Since EEH triggers are an exceedingly corner case error condition,
> this puts this patch even less valid to go in outside of the merge
> window.
Just in case it's lack of testing you were concerned with, I can say 
that I've tested this patch on power and it does fix our eeh handling 
for bnx2x.

Thanks,
Paul Larson