Re: USB-C device hotplug issue

[Dennis Wassenberg <dennis.wassenberg@secunet.com>]

On 09.11.18 14:47, Mathias Nyman wrote:
> On 07.11.2018 11:08, Dennis Wassenberg wrote:
>>
>> On 05.11.18 16:35, Mathias Nyman wrote:
>>> On 26.10.2018 17:07, Alan Stern wrote:
>>>> On Fri, 26 Oct 2018, Dennis Wassenberg wrote:
>>>>
>>>>>>> --- a/drivers/usb/core/hub.c
>>>>>>> +++ b/drivers/usb/core/hub.c
>>>>>>> @@ -2815,7 +2815,9 @@ static int hub_port_reset(struct usb_hub *hub, int port1,
>>>>>>>                        USB_PORT_FEAT_C_BH_PORT_RESET);
>>>>>>>                usb_clear_port_feature(hub->hdev, port1,
>>>>>>>                        USB_PORT_FEAT_C_PORT_LINK_STATE);
>>>>>>> -            usb_clear_port_feature(hub->hdev, port1,
>>>>>>> +
>>>>>>> +            if (!warm)
>>>>>>> +                usb_clear_port_feature(hub->hdev, port1,
>>>>>>>                        USB_PORT_FEAT_C_CONNECTION);
>>>>>>>                  /*
>>>>>>
>>>>>> The key fact is that connection events can get lost if they happen to
>>>>>> occur during a port reset.
>>>>> Yes.
>>>>>>
>>>>>> I'm not entirely certain of the logic here, but it looks like the
>>>>>> correct test to add should be "if (udev != NULL)", not "if (!warm)".
>>>>>> Perhaps Mathias can confirm this
>>>
>>> Sorry about the late response, got distracted while performing git
>>> archeology.
>>>
>>> "if (udev != NULL)" would seem more reasonable.
>>>
>>> Logs show that clearing the FEAT_C_CONNECTION was originally added
>>> after a hot reset failed, and before issuing a warm reset to a SS.Inactive
>>> link.  (see 10d674a USB: When hot reset for USB3 fails, try warm reset.)
>>>
>>> Apparently all the change flags needed to be cleared for some specific
>>> host + device combination before issuing a warm reset for the enumeration
>>> to work properly.
>>>
>>> The change to always clear FEAT_C_CONNECTION for USB3 was done later in patch:
>>> a24a607 USB: Rip out recursive call on warm port reset.
>>>
>>> Motivation was:
>>>
>>> "In hub_port_finish_reset, unconditionally clear the connect status
>>>   change (CSC) bit for USB 3.0 hubs when the port reset is done.  If we
>>>   had to issue multiple warm resets for a device, that bit may have been
>>>   set if the device went into SS.Inactive and then was successfully warm
>>>   reset."
>>>
>>>>> I don't know if clearing the USB_PORT_FEAT_C_CONNECTION bit is
>>>>> correct in case of a non warm reset. In my case I always observed a
>>>>> warm reset because of the link state change.
>>>>> Thats why I checked the warm variable to not change the behavoir for
>>>>> cases a didn't checked for the first shot.
>>>>
>>>> (The syntax of that last sentence is pretty mangled; I can't understand
>>>> it.)
>>>>
>>>> Think of it this way:
>>>>
>>>>      If a device was not attached before the reset, we don't want
>>>>      to clear the connect-change status because then we wouldn't
>>>>      recognize a device that was plugged in during the reset.
>>>>
>>>>      If a device was attached before the reset, we don't want any
>>>>      connect-change status which might be provoked by the reset to
>>>>      last, because we don't want the core to think that the device
>>>>      was unplugged and replugged when all that happened was a reset.
>>>>
>>>> So the important criterion is whether or not a device was attached to
>>>> the port when the reset started.  It's something of a coincidence that
>>>> you only observe warm resets when there's nothing attached.
>>>>
>>>>> During the first run of usb_hub_reset the udev is NULL because in
>>>>> this situation the device is not attached logically.
>>>>>
>>>>> [  112.889810] usb 4-1-port1: XXX: port_event: portstatus: 0x2c0, portchange: 0x40!
>>>>> [  113.201192] usb 4-1-port1: XXX: hub_port_reset: udev:            (nil)!
>>>>> [  113.201198] usb 4-1-port1: XXX: hub_port_reset (not clearing USB_PORT_FEAT_C_CONNECTION): 0x203, portchange: 0x1!
>>>>> [  113.253612] usb 4-1-port1: XXX: port_event: portstatus: 0x203, portchange: 0x1!
>>>>> [  113.377208] usb 4-1-port1: XXX: hub_port_reset: udev: ffff88046b302800!
>>>>> [  113.377214] usb 4-1-port1: XXX: hub_port_reset (not clearing USB_PORT_FEAT_C_CONNECTION): 0x203, portchange: 0x0!
>>>>> [  113.429478] usb 4-1.1: new SuperSpeed USB device number 7 using xhci_hcd
>>>>> [  113.442370] usb 4-1.1: New USB device found, idVendor=0781, idProduct=5596
>>>>> [  113.442376] usb 4-1.1: New USB device strings: Mfr=1, Product=2, SerialNumber=3
>>>>> [  113.442381] usb 4-1.1: Product: Ultra T C
>>>>> [  113.442385] usb 4-1.1: Manufacturer: SanDisk
>>>>> [  113.442388] usb 4-1.1: SerialNumber: 4C530001131013121031
>>>>>
>>>>> Or maybe we can skip clearing the USB_PORT_FEAT_C_CONNECTION bit in
>>>>> case of hub_port_reset completely without any other check?
>>>>
>>>> See above.
>>>
>>> Checking for udev sounds reasonable to me.
>>> Not sure if we should worry about the specific host+device combo that needed flags
>>> cleared before warm reset. See patch:
>>>
>>> 10d674a USB: When hot reset for USB3 fails, try warm reset.
>>> https://marc.info/?l=linux-usb&m=131603549603799&w=2
>>>
>>> -Mathias
>> Checking for udev works well too in my case. So there is no need to check the warm flag.
>>
>> Regarding the other point about the specific host+device combo which needs the flags cleared, I don't know how to
>> proceed.
>>
> 
> I support just adding a udev check patch, want to send one?
Ok, I will do so.
> 
> Current hub port reset code is wrong, causing real life issues today.
> 
> The issue with the specific host+device is from 2011,
> The port reset code has changed completely since then.
> If it turns out to still be a issue we can make a host/device specific quirk.
>> -Mathias
ok, understood.

Dennis