On Wed, Apr 24, 2024 at 03:33:39AM +0000, Smita Koralahalli wrote:
> Clear Link Bandwidth Management Status (LBMS) if set, on a hot-remove event.
>
> The hot-remove event could result in target link speed reduction if LBMS
> is set, due to a delay in Presence Detect State Change (PDSC) happening
> after a Data Link Layer State Change event (DLLSC).
>
> In reality, PDSC and DLLSC events rarely come in simultaneously. Delay in
> PDSC can sometimes be too late and the slot could have already been
> powered down just by a DLLSC event. And the delayed PDSC could falsely be
> interpreted as an interrupt raised to turn the slot on. This false process
> of powering the slot on, without a link forces the kernel to retrain the
> link if LBMS is set, to a lower speed to restablish the link thereby
> bringing down the link speeds [2].
>
> According to PCIe r6.2 sec 7.5.3.8 [1], it is derived that, LBMS cannot
> be set for an unconnected link and if set, it serves the purpose of
> indicating that there is actually a device down an inactive link.
> However, hardware could have already set LBMS when the device was
> connected to the port i.e when the state was DL_Up or DL_Active. Some
> hardwares would have even attempted retrain going into recovery mode,
> just before transitioning to DL_Down.
>
> Thus the set LBMS is never cleared and might force software to cause link
> speed drops when there is no link [2].
Is this an issue introduced by commit a89c82249c37 ("PCI: Work around
PCIe link training failures")?
If so, could you add a Fixes tag?
I might be mistaken but my impression is that we're seeing a lot of
fallout as a result of that commit.
Thanks,
Lukas
