tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: Multiple outstanding transfer vs xhci / ehci (Re: CVS commit: src/sys/dev/usb)



On 2019/02/15 23:37, Nick Hudson wrote:
On 15/02/2019 13:44, Rin Okuyama wrote:
On 2019/02/15 21:57, Jonathan A. Kollasch wrote:
On Wed, Feb 13, 2019 at 06:35:14PM +0900, Rin Okuyama wrote:
Hi,

On 2019/02/13 3:54, Nick Hudson wrote:
On 12/02/2019 16:02, Rin Okuyama wrote:
Hi,

The system freezes indefinitely with xhci(4) or ehci(4), when NIC with
multiple outstanding transfers [axen(4), mue(4), and ure(4)] is stopped
by "ifconfig down" or detached.

As discussed in the previous message, this is due to infinite loop in
usbd_ar_pipe(); xfers with USBD_NOT_STARTED remain in a queue forever
because upm_abort [= xhci_device_bulk_abort() etc.] cannot remove them.


Why not the attached patch instead?

Nick

Thank you so much for your prompt reply!

It works if s/ux_state/ux_status/ here:

+        if (xfer->ux_state == USBD_NOT_STARTED) {

Can I commit the revised patch?


The revised patch results in
https://nxr.netbsd.org/xref/src/sys/dev/usb/ehci.c?r=1.265#1566
firing upon `drvctl detach -d bwfm0` on Pinebook.

    Jonathan Kollasch

IMO, this is because NOT_STARTED queues are removed in a way that is
unexpected to ehci. For my old amd64 box, the attached patch fixes
assertion failures when devices are detached from ehci. I would like
to commit it together with the previous patch.


ux_state has probably out lived its usefulness.

Other/All HCs do the same ux_state check. So, either all need to be updated or we do the same XFER_ONQU to XFER_BUSY transition in usbd_ar_pipe


Nick

The latter does not work because ehci and uhci also check its own flags,
ex_isdone and ux_isdone, respectively. Therefore, I chose the former:

http://www.netbsd.org/~rin/usbhc_freex_20190215.patch

Is this OK for you?

Thanks,
rin


Home | Main Index | Thread Index | Old Index