[net-next-2.6 PATCH] e1000e: don't inadvertently re-set INTX_DISABLE

[net-next-2.6 PATCH] e1000e: don't inadvertently re-set INTX_DISABLE

[Dean Nelson]

Should e1000_test_msi() fail to see an msi interrupt, it attempts to
fallback to legacy INTx interrupts. But an error in the code may prevent
this from happening correctly.

Before calling e1000_test_msi_interrupt(), e1000_test_msi() disables SERR
by clearing the SERR bit from the just read PCI_COMMAND bits as it writes
them back out.

Upon return from calling e1000_test_msi_interrupt(), it re-enables SERR
by writing out the version of PCI_COMMAND it had previously read.

The problem with this is that e1000_test_msi_interrupt() calls
pci_disable_msi(), which eventually ends up in pci_intx(). And because
pci_intx() was called with enable set to 1, the INTX_DISABLE bit gets
cleared from PCI_COMMAND, which is what we want. But when we get back to
e1000_test_msi(), the INTX_DISABLE bit gets inadvertently re-set because
of the attempt by e1000_test_msi() to re-enable SERR.

The solution is to have e1000_test_msi() re-read the PCI_COMMAND bits as
part of its attempt to re-enable SERR.

During debugging/testing of this issue I found that not all the systems
I ran on had the SERR bit set to begin with. And on some of the systems
the same could be said for the INTX_DISABLE bit. Needless to say these
latter systems didn't have a problem falling back to legacy INTx
interrupts with the code as is.

Signed-off-by: Dean Nelson <dnelson@redhat.com>
CC: stable@kernel.org

---

I forced the msi interrupt test to fail by commenting out the clearing
of FLAG_MSI_TEST_FAILED in e1000_intr_msi_test().

 drivers/net/e1000e/netdev.c |   13 +++++++++----
 1 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/drivers/net/e1000e/netdev.c b/drivers/net/e1000e/netdev.c
index 57a7e41..79e38dc 100644
--- a/drivers/net/e1000e/netdev.c
+++ b/drivers/net/e1000e/netdev.c
@@ -3419,13 +3419,18 @@ static int e1000_test_msi(struct e1000_adapter *adapter)
 
 	/* disable SERR in case the MSI write causes a master abort */
 	pci_read_config_word(adapter->pdev, PCI_COMMAND, &pci_cmd);
-	pci_write_config_word(adapter->pdev, PCI_COMMAND,
-			      pci_cmd & ~PCI_COMMAND_SERR);
+	if (pci_cmd & PCI_COMMAND_SERR)
+		pci_write_config_word(adapter->pdev, PCI_COMMAND,
+				      pci_cmd & ~PCI_COMMAND_SERR);
 
 	err = e1000_test_msi_interrupt(adapter);
 
-	/* restore previous setting of command word */
-	pci_write_config_word(adapter->pdev, PCI_COMMAND, pci_cmd);
+	/* re-enable SERR */
+	if (pci_cmd & PCI_COMMAND_SERR) {
+		pci_read_config_word(adapter->pdev, PCI_COMMAND, &pci_cmd);
+		pci_cmd |= PCI_COMMAND_SERR;
+		pci_write_config_word(adapter->pdev, PCI_COMMAND, pci_cmd);
+	}
 
 	/* success ! */
 	if (!err)

Re: [net-next-2.6 PATCH] e1000e: don't inadvertently re-set INTX_DISABLE

[Jeff Kirsher]

On Fri, Jun 18, 2010 at 13:36, Dean Nelson <dnelson@redhat.com> wrote:
> Should e1000_test_msi() fail to see an msi interrupt, it attempts to
> fallback to legacy INTx interrupts. But an error in the code may prevent
> this from happening correctly.
>
> Before calling e1000_test_msi_interrupt(), e1000_test_msi() disables SERR
> by clearing the SERR bit from the just read PCI_COMMAND bits as it writes
> them back out.
>
> Upon return from calling e1000_test_msi_interrupt(), it re-enables SERR
> by writing out the version of PCI_COMMAND it had previously read.
>
> The problem with this is that e1000_test_msi_interrupt() calls
> pci_disable_msi(), which eventually ends up in pci_intx(). And because
> pci_intx() was called with enable set to 1, the INTX_DISABLE bit gets
> cleared from PCI_COMMAND, which is what we want. But when we get back to
> e1000_test_msi(), the INTX_DISABLE bit gets inadvertently re-set because
> of the attempt by e1000_test_msi() to re-enable SERR.
>
> The solution is to have e1000_test_msi() re-read the PCI_COMMAND bits as
> part of its attempt to re-enable SERR.
>
> During debugging/testing of this issue I found that not all the systems
> I ran on had the SERR bit set to begin with. And on some of the systems
> the same could be said for the INTX_DISABLE bit. Needless to say these
> latter systems didn't have a problem falling back to legacy INTx
> interrupts with the code as is.
>
> Signed-off-by: Dean Nelson <dnelson@redhat.com>
> CC: stable@kernel.org
>
> ---
>
> I forced the msi interrupt test to fail by commenting out the clearing
> of FLAG_MSI_TEST_FAILED in e1000_intr_msi_test().
>
>  drivers/net/e1000e/netdev.c |   13 +++++++++----
>  1 files changed, 9 insertions(+), 4 deletions(-)
>

Thanks Dean, I have added it to my queue.

Cheers,
Jeff