On Tue, 2018-11-13 at 10:36 +0000, Lorenzo Pieralisi wrote:
On Tue, Nov 13, 2018 at 01:18:41AM +0000, Trent Piepho wrote:
[...]
Based on the information provided, its very likely that I should use the PCIE_MSI_INTR0_MASK register instead of PCIE_MSI_INTR0_ENABLE on the dw_pci_bottom_mask() and dw_pci_bottom_unmask() callbacks. As soon as I return from Linux plumbers conference, I will test the system using the PCIE_MSI_INTR0_MASK register.
Using enable to mask the interrupt is broken. It will allow an MSI to be lost if it arrives while the MSI in not enabled. It is impossible to prevent that from racing against the pci device driver's final check that no MSI-signaled condition in the hardware is present.
That's exactly why I raised the point. I would like to understand if this means that this revert is not enough to fix the underlying issue or we need this revert AND correct mask/unmask operations.
Looks like _another_ bug added to the driver, this time in 4.17.