On Wed, 11 Dec 2013, Julius Werner wrote: > >> ...although, the spec says that it does not wait for the port resets > >> to complete. As far as I can see re-issuing a warm reset and waiting > >> is the only way to guarantee the core times the recovery. Presumably > >> the portstatus debounce in hub_activate() mitigates this, but that > >> 100ms is less than a full reset timeout. > > It's definitely not just a timing issue for us. I can't reproduce all > the same cases as Vikas, but when I attach a USB analyzer to the ones > I do see the host controller doesn't even start sending a reset. > > >>> The xHCI spec requires that when the xHCI host is reset, a USB reset is > >>> driven down the USB 3.0 ports. If hot reset fails, the port may migrate > >>> to warm reset. See table 32 in the xHCI spec, in the definition of > >>> HCRST. It sounds like this host doesn't drive a USB reset down USB 3.0 > >>> ports at all on host controller reset? > > Oh, interesting, I hadn't seen that yet. So I guess the spec itself is > fine if it were followed to the letter. > > I did some more tests about this on my Exynos machine: when I put a > device to autosuspend (U3) and manually poke the xHC reset bit, I do > see an automatic warm reset on the analyzer and the ports manage to > retrain to U0. But after a system suspend/resume which calls > xhci_reset() in the process, there is no reset on the wire. I also > noticed that it doesn't drive a reset (even after manual poking) when > there is no device connected on the other end of the analyzer. > > So this might be our problem: maybe these host controllers (Synopsys > DesignWare) issue the spec-mandated warm reset only on ports where > they think there is a device attached. But after a system > suspend/resume (where the whole IP block on the SoC was powered down), > the host controller cannot know that there is still a device with an > active power session attached, and therefore doesn't drive the reset > on its own. > > Even though this is a host controller bug, we still have to deal with > it somehow. I guess we could move the code into xhci_plat_resume() and > hide it behind a quirk to lessen the impact. But since reset_resume is > not a common case for most host controllers, it's hard to say if this > is DesignWare specific or a more widespread implementation mistake.
I was going to suggest something along these lines too. This seems to be a bug in xHCI. Therefore the fix belongs in xhci-hcd, not in the hub driver. Alan Stern -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
