On Fri, 10 Jan 2025, Jiwei Sun wrote: > In order to fix the issue, don't do the retraining work except ASMedia > ASM2824. I yet need to go through all of your submission in detail, but this assumption defeats the purpose of the workaround, as the current understanding of the origin of the training failure and the reason to retrain by hand with the speed limited to 2.5GT/s is the *downstream* device rather than the ASMedia ASM2824 switch. It is also why the quirk has been wired to run everywhere rather than having been keyed by VID:DID, and the VID:DID of the switch is only listed, conservatively, because it seems safe with the switch to lift the speed restriction once the link has successfully completed training. Overall I think we need to get your problem sorted differently, because I suppose in principle your hot-plug scenario could also happen with the ASMedia ASM2824 switch as the upstream device and your NVMe storage element as the downstream device. Perhaps the speed restriction could be always lifted, and then the bandwidth controller infrastructure used for that, so that it doesn't have to happen within `pcie_failed_link_retrain'? Maciej
