e1000e: Fixed issue with LSC pending.
Subsystems (2)
PATCH 1/1 e1000e: Fixed issue with LSC pending.
From: Andrew Melnychenko <[email protected]>
The issue was detected with QEMU and doesn't reproduce on Windows guest. CTRL_EXT.IAME is disabled by
rev 0a8047ac68e50e4ccbadcfc6b6b070805b976885:
"to avoid disruption from potential programs that access the registers directly."
So let's leave it like that and add interrupt pending clearance using write to ICR. It fixes the issue when removed cable can't be detected by the driver.
--- drivers/net/ethernet/intel/e1000e/netdev.c | 3 +++ 1 file changed, 3 insertions(+)
drivers/net/ethernet/intel/e1000e/netdev.c +3 -0
@@ -1898,6 +1898,9 @@ static irqreturn_t e1000_msix_other(int __always_unused irq, void *data)struct e1000_hw *hw = &adapter->hw;u32 icr =er32 (ICR ); +if (!(er32 (CTRL_EXT ) &E1000_CTRL_EXT_IAME )) +ew32 (ICR , icr & ~E1000_ICR_INT_ASSERTED ); +if (icr & adapter->eiac_mask)ew32 (ICS , (icr & adapter->eiac_mask));
Signature · 1 line
2.24.1
PATCH 1/1 e1000e: Fixed issue with LSC pending.
Quoted 11 lines · On Fri, 17 Apr 2020 14:06:27 +0300 <[email protected]> wrote:
From: Andrew Melnychenko <[email protected]>
The issue was detected with QEMU and doesn't reproduce on Windows guest. CTRL_EXT.IAME is disabled by
rev0a8047ac68e50e4ccbadcfc6b6b070805b976885:
"to avoid disruption from potential programs that access the registers directly."
So let's leave it like that and add interrupt pending clearance using write to ICR. It fixes the issue when removed cable can't be detected by the driver.
Thanks for your patch! I think you may want to add a correctly formatted Fixes: tag.
Please beware, this interrupt code is very tricky and difficult to get right, we had a "small" change in e1000 that took months to iron out and get working correctly, and in the meantime broke a bunch of stuff.
So, explaining root cause in your commit message, and a comment below to help future code editors understand the code/functionality better would be a good idea, IMO. Also, the QEMU behavior often doesn't do a very good job of matching real hardware (esp with side effect registers like ICR), so sometimes it is better to just fix QEMU, did you consider that in this case?
Quoted 11 lines
Signed-off-by Andrew Melnychenko <[email protected]>--- drivers/net/ethernet/intel/e1000e/netdev.c | 3 +++ 1 file changed, 3 insertions(+)drivers/net/ethernet/intel/e1000e/netdev.c +0
-0
drivers/net/ethernet/intel/e1000e/netdev.c static irqreturn_t e1000_msix_other(int __always_unused irq, void *data) · line 1900
Quoted 3 lines
drivers/net/ethernet/intel/e1000e/netdev.c +0
-0@@ -1898,6 +1898,9 @@ static irqreturn_t e1000_msix_other(int __always_unused irq, void *data)struct e1000_hw *hw = &adapter->hw;u32 icr =er32 (ICR );
I'd really like to see a comment in the code here explaining what is going on. f.e.
/* if auto-mask is not enabled we need to explicitly * clear the interrupt asserted bit in ICR to allow * more interrupts */
Quoted 5 lines
drivers/net/ethernet/intel/e1000e/netdev.c +3
-0+if (!(er32 (CTRL_EXT ) &E1000_CTRL_EXT_IAME )) +ew32 (ICR , icr & ~E1000_ICR_INT_ASSERTED ); +if (icr & adapter->eiac_mask)ew32 (ICS , (icr & adapter->eiac_mask));
PATCH 1/1 e1000e: Fixed issue with LSC pending.
In reply to [email protected]Quoted 23 lines · On Fri, Apr 17, 2020 at 3:39 AM <[email protected]> wrote:
From: Andrew Melnychenko <[email protected]>
The issue was detected with QEMU and doesn't reproduce on Windows guest. CTRL_EXT.IAME is disabled by
rev0a8047ac68e50e4ccbadcfc6b6b070805b976885:
"to avoid disruption from potential programs that access the registers directly."
So let's leave it like that and add interrupt pending clearance using write to ICR. It fixes the issue when removed cable can't be detected by the driver.Signed-off-by Andrew Melnychenko <[email protected]>--- drivers/net/ethernet/intel/e1000e/netdev.c | 3 +++ 1 file changed, 3 insertions(+)drivers/net/ethernet/intel/e1000e/netdev.c +0
-0
drivers/net/ethernet/intel/e1000e/netdev.c static irqreturn_t e1000_msix_other(int __always_unused irq, void *data) · line 1906
Quoted 9 lines
drivers/net/ethernet/intel/e1000e/netdev.c +3
-0@@ -1898,6 +1898,9 @@ static irqreturn_t e1000_msix_other(int __always_unused irq, void *data)struct e1000_hw *hw = &adapter->hw;u32 icr =er32 (ICR ); +if (!(er32 (CTRL_EXT ) &E1000_CTRL_EXT_IAME )) +ew32 (ICR , icr & ~E1000_ICR_INT_ASSERTED ); +if (icr & adapter->eiac_mask)ew32 (ICS , (icr & adapter->eiac_mask));
So this would likely cause more issues then it helps. Specifically, clearing the ICR should happen with any read of it as long as at least one bit in the current mask is set, and the write doesn't make any sense for actual hardware as the INT_ASSERTED bit is not clear on write. It is cleared when the register is cleared.
I am pretty sure this is a bug in the QEMU side of the implementation, specifically in e1000e_mac_icr_read
(https://lxr.missinglinkelectronics.com/qemu+v4.0.0/hw/net/e1000e_core.c#L2608).
If you look at section 13.3.27 of the 8257X developers manual
(https://lxr.missinglinkelectronics.com/qemu+v4.0.0/hw/net/e1000e_core.c#L2608)
it has a lengthy explanation about the ICR register in the "Note:" section where it essentially explains that as long as any bit in the IMS is set that corresponds to the ICR register value, and the INT_ASSERTED bit is set then the entire ICR is cleared on read. The IAME bit triggers something similar but it is based on the IAM register which is why it worked prior to this patch.
The fix would likely to be to add a bit to QEMU to allow for clearing the ICR register, or at least any bits from it that are set in the IMS when it is read.