mbox series

[RFC,0/3] Fix enumeration issues during composition switching

Message ID 20220203080017.27339-1-quic_wcheng@quicinc.com (mailing list archive)
Headers show
Series Fix enumeration issues during composition switching | expand

Message

Wesley Cheng Feb. 3, 2022, 8 a.m. UTC
This patch series addresses a few enumeration issues being seen during fast
composition switching test cases:
  - Missing DWC3 core soft reset before run/stop enable.  Recommended by the
    Synopsis databook to do this during the pullup enable case.
  - Endxfer command timeouts leading to controller halt failures in the pullup
    disable path.
  - Endxfer command timeouts leading to EP dequeue to never return a completion
    to the function drivers, which can cause to unbind issues depending on
    function driver implementation.

With regards to the endxfer timeout, it was communicated to us that if there is
a pending setup/control transfer in progress, the DWC3 controller is unable to
service the endxfer command for other endpoints.  USB bus sniffer and USB ftrace
logs confirmed that when the endxfer timeouts occurred, there was a SETUP token
being sent by the host, while the pullup routine was in progress, as during the
pullup disable, we should not see any ftrace logs since IRQs are disabled.

It was recommended by Synopsis to ensure that packets on EP0 are always handled,
and the only way to drain the dedicated control transfer internal memory is to
issue a startxfer command.  With this in mind, the pullup disable case is able
to discard any cached setup packets (as disconnect would pursue), so the
approach to issue a startxfer after an unsuccessful endxfer is possible.
Test logs show the subsequent endxfer is successful:

[004] d..1 15631.849982: dwc3_gadget_ep_cmd: ep1out: cmd 'End Transfer' [20c08] params 00000000 00000000 00000000 --> status: Timed Out
[004] d..1 15631.850008: dwc3_prepare_trb: ep0out: trb ffffffc019dad000 (E0:D0) buf 00000000efffa000 size 8 ctrl 00000c23 (HLcs:SC:setup)
[004] d..1 15631.850024: dwc3_gadget_ep_cmd: ep0out: cmd 'Start Transfer' [406] params 00000000 efffa000 00000000 --> status: Successful
[004] d..1 15631.857380: dwc3_gadget_ep_cmd: ep1out: cmd 'End Transfer' [20c08] params 00000000 00000000 00000000 --> status: Successful
[004] d..1 15631.857409: dwc3_gadget_giveback: ep1out: req ffffff8789212300 length 0/16384 zsI ==> -108

Attempts to ensure EP0 transfers don't happen during pullup disable were added
as well.

For the dequeue path, since usb_ep_dequeue() can be called from the function
driver at any time, the same approach can not be used.  In this case, if there
is an endxfer failure observed when dequeuing a request, mark the endpoint w/ a 
flag, which will be later checked when the control transfer is complete.  From
there it will traverse through all EPs to service the ones that need to
re-issue the endxfer command.  This logic will only trigger if there is at least
one EP that needs servicing.

Wesley Cheng (3):
  usb: dwc3: Flush pending SETUP data during stop active xfers
  usb: dwc3: gadget: Wait for ep0 xfers to complete during dequeue
  usb: dwc3: Issue core soft reset before enabling run/stop

 drivers/usb/dwc3/core.c   |   4 +-
 drivers/usb/dwc3/core.h   |  14 ++++++
 drivers/usb/dwc3/ep0.c    |  23 ++++++---
 drivers/usb/dwc3/gadget.c | 100 +++++++++++++++++++++++++++++++++-----
 4 files changed, 118 insertions(+), 23 deletions(-)