[v8,0/5] Fix controller halt and endxfer timeout issues

Message

Wesley Cheng Sept. 1, 2022, 7:36 p.m. UTC

Changes in v8:
- Remove stub for dwc3_remove_requests() as if DWC3 host only is enabled
then ep0 isn't event compiled.

Changes in v7:
- Added missing arugment if DWC3 is not enabled in the kernel config

Changes in v6:
- Fix kernel bot errors/warnings.

Changes in v5:
- Rebased series on usb-testing from patch #5 onwards.

Changes in v4:
- Split the increase timeout patch into separate patches. #1 for the
gadget suspend/resume locking changes #2 for the increased timeout
- Modified msleep to usleep_range w/ an interval of 1-2ms and a max
timeout of 4s.

Changes in v3:
- Modified the msleep() duration to ~2s versus ~10s due to the minimum
mdelay() value.
- Removed patch to modify DEP flags during dwc3_stop_active_transfer().
This was not required after fixing the logic to allow EP xfercomplete
events to be handled on EP0.
- Added some changes to account for a cable disconnect scenario, where
dwc3_gadget_pullup() would not be executed to stop active transfers.
Needed to add some logic to the disconnect interrupt to ensure that we
cleanup/restart any pending SETUP transaction, so that we can clear the
EP0 delayed stop status. (if pending)
- Added patch to ensure that we don't proceed with umapping buffers
until the endxfer was actually sent.

Changes in v2:
- Moved msleep() to before reading status register for halted state
- Fixed kernel bot errors
- Clearing DEP flags in __dwc3_stop_active_transfers()
- Added Suggested-by tags and link references to previous discussions

This patch series addresses some issues seen while testing with the latest
soft disconnect implementation where EP events are allowed to process while
the controller halt is occurring.

#1
Since routines can now interweave, we can see that the soft disconnect can
occur while conndone is being serviced.  This leads to a controller halt
timeout, as the soft disconnect clears the DEP flags, for which conndone
interrupt handler will issue a __dwc3_ep_enable(ep0), that leads to
re-issuing the set ep config command for every endpoint.

#2
Function drivers can ask for a delayed_status phase, while it processes the
received SETUP packet.  This can lead to large delays when handling the
soft disconnect routine.  To improve the timing, forcefully send the status
phase, as we are going to disconnect from the host.

#3
Ensure that local interrupts are left enabled, so that EP0 events can be
processed while the soft disconnect/dequeue is happening.

#4
Since EP0 events can occur during controller halt, it may increase the time
needed for the controller to fully stop.

#5
Account for cable disconnect scenarios where nothing may cause the endxfer
retry if DWC3_EP_DELAY_STOP is set.

#6
Avoid unmapping pending USB requests that were never stopped.  This would
lead to a potential SMMU fault.


Wesley Cheng (5):
  usb: dwc3: Avoid unmapping USB requests if endxfer is not complete
  usb: dwc3: Remove DWC3 locking during gadget suspend/resume
  usb: dwc3: Increase DWC3 controller halt timeout
  usb: dwc3: gadget: Skip waiting for CMDACT cleared during endxfer
  usb: dwc3: gadget: Submit endxfer command if delayed during disconnect

 drivers/usb/dwc3/core.c   |  4 ----
 drivers/usb/dwc3/core.h   |  1 +
 drivers/usb/dwc3/ep0.c    |  5 ++++-
 drivers/usb/dwc3/gadget.c | 31 +++++++++++++++++++++++++++----
 4 files changed, 32 insertions(+), 9 deletions(-)

Comments

Greg Kroah-Hartman Sept. 7, 2022, 2:27 p.m. UTC | #1

On Thu, Sep 01, 2022 at 12:36:20PM -0700, Wesley Cheng wrote:
> Changes in v8:
> - Remove stub for dwc3_remove_requests() as if DWC3 host only is enabled
> then ep0 isn't event compiled.

This only applied with manual override of some fuzz in patch 2/5, which
is odd.  You should check that it all applied correctly to my tree
please.

thanks,

greg k-h