On Sat, Mar 11, 2023 at 09:22:20AM +0100, Lukas Wunner wrote: > > - Switch and NVMe MPS are 512B > > - NVMe config space saved (including MPS=512B) > > - You change Switch MPS to 128B > > - NVMe does DMA with payload > 128B > > - Switch reports Malformed TLP because TLP is larger than its MPS > > - Recovery resets NVMe, which sets MPS to the default of 128B > > - nvme_slot_reset() restores NVMe config space (MPS is now 512B) > > - Subsequent NVMe DMA with payload > 128B repeats cycle > > Forgive my ignorance, but if MPS is restored to 512B by nvme_slot_reset(), > shouldn't the communication with the device just work again from that > point on? The upstream port was tuned down 128 without coordinating with the kernel, so restoring the nvme to 512 creates the mismatch.
