Re: Missing USB XHCI and EHCI reset for kexec

[Stefani Seibold <stefani@seibold.net>]

Am Dienstag, den 15.04.2014, 15:33 -0300 schrieb Thadeu Lima de Souza
Cascardo:
> On Tue, Apr 15, 2014 at 05:00:28PM +0200, stefani@seibold.net wrote:
> > 
> > Zitat von Thadeu Lima de Souza Cascardo <cascardo@linux.vnet.ibm.com>:
> > 
> > >On Tue, Apr 15, 2014 at 12:04:17PM +0200, stefani@seibold.net wrote:
> > >>
> > >>Zitat von Thadeu Lima de Souza Cascardo <cascardo@linux.vnet.ibm.com>:
> > >>
> > >>>On Mon, Apr 14, 2014 at 05:44:58PM +0200, stefani@seibold.net wrote:
> > >>>>
> > >>>>Zitat von Benjamin Herrenschmidt <benh@au1.ibm.com>:
> > >>>>
> > >>>>>I don't know about EHCI specifically but this is a known issue with
> > >>>>>XHCI, I observe similar issues on other powerpc platforms (servers)
> > >>>>>and this isn't architecture specific (looks more like actualy xhc
> > >>>>>implementation specific).
> > >>>>>
> > >>>>>Thadeu Cascardo (on CC) has been the one investigating that on our side,
> > >>>>>he might have more to add including patches.
> > >>>>>
> > >>>>
> > >>>>I have now a kernel 3.14 dmesg log of the problem. After a kexec the
> > >>>>kexeced 3.14 kernel shows:
> > >>>>
> > >>>>[    1.170029] xhci_hcd 0001:03:00.0: xHCI Host Controller
> > >>>>[    1.175306] xhci_hcd 0001:03:00.0: new USB bus registered,
> > >>>>assigned bus number 1
> > >>>>[    1.212561] xhci_hcd 0001:03:00.0: Host not halted after 16000
> > >>>>microseconds.
> > >>>>[    1.219621] xhci_hcd 0001:03:00.0: can't setup: -110
> > >>>>[    1.224597] xhci_hcd 0001:03:00.0: USB bus 1 deregistered
> > >>>>[    1.230021] xhci_hcd 0001:03:00.0: init 0001:03:00.0 fail, -110
> > >>>>[    1.235955] xhci_hcd: probe of 0001:03:00.0 failed with error -110
> > >>>>
> > >>>
> > >>>What is your controller vendor and device IDs? Is that a TI chip?
> > >>>
> > >>
> > >>Yes it is a TI chip, vendor ID 104c and product ID 8241.
> > >>
> > >>>Can you check if the patch I sent a month ago fixes it? [1] There's the
> > >>>whole story there. In fact, you will also need something like the patch
> > >>>below. Can you apply only the first one, verify, and, then, the other
> > >>>one as well, and report what worked for you?
> > >>>
> > >>>[1] http://marc.info/?l=linux-usb&m=139483181809062&w=2
> > >>>
> > >>
> > >>I tried the attach patch and it did not help. This is what i
> > >>expected because this is a fix in the shutdown path, which will
> > >>never called when doing a forced kexec.
> > >
> > >Hi, Stefani.
> > >
> > >Did you try with both patches applied? How do you evoke the forced
> > >kexec? Is that a kexec on panic? Does it really need to be forced? With
> > >no clean shutdown, platform and drivers would need to issue resets, like
> > >you mentioned below, to get the system into a clean state.
> > >
> > 
> > Yes, i applied both patches. But without success.
> > 
> > IMHO i think it is necessary to bring the device i a clean state
> > when the driver use the HW.
> > 
> > >>
> > >>I have a running a 3.10.23 kernel. This kernel do a kexec for a
> > >>kernel 3.14. Since the kernel 3.10.23 did not performe a clean
> > >>shutdown, the state of the XHCI Controller is undefined. So when
> > >
> > >And the clean shutdown requires both of my patches, for TI chips, as far
> > >as I know. It looks like the problem is issuing a halt when there are
> > >pending URBs.
> > >
> > >>kernel 3.14 will probe XHCI it will find a XHCI controller which was
> > >>not performed a reset.
> > >>
> > >
> > >The problem is not that a reset hasn't been issued. A PCI function reset
> > >should fix most of the problems with a bad device state, when the reset
> > >works. However, the problem is that it was not cleanly shut down. URBs
> > >should have been canceled and removed from the controller queue, and it
> > >should have halted after that.
> > 
> > Again, i think it is the job of the driver to bring the chip in a clean state
> > before using them. A driver should never expect a reset state of a chip.
> > 
> > >
> > >>So i think it is necessary to reset the XHCI controller and all
> > >>devices on this bus. This is what i do with a "echo 1
> > >>>/sys/bus/pci/drivers/xhci_hcd/0001:03:00.0/reset" before the kexec.
> > >>
> > >
> > >One way to look at that is making the PCI code issue resets to all buses
> > >before doing any other access. That will make booting more slow, and
> > >there are a lot of other corner cases where this might not be enough.
> > >It's probably more sane to try to get the 3.10.23 kernel to do a clean
> > >shutdown, if possible.
> > >
> > 
> > With this driver design the kexec functionality is usesless on PowerPC.
> > X86 looks a little bit better.
> > 
> > - Stefani
> > 
> > 
> 
> What is the vendor and device ID you are using on your X86 system? This
> is not a matter of what architecture you are using, it's the XHCI
> controller which does not behave as well as the one you are using on
> X86, which is likely an Intel one.
> 

It is an Intel 8086:8c31. But this was only a side note. We need a
generic solution not a vendor specific one. Otherwise kexec is useless
on other architectures.

- Stefani