Re: [PATCH 4/5] usb: dwc3: Allow end transfer commands to be sent during soft disconnect

From: Wesley Cheng
Date: Wed Jul 13 2022 - 20:33:52 EST

Hi Thinh,

On 7/13/2022 5:19 PM, Thinh Nguyen wrote:
On 7/13/2022, Wesley Cheng wrote:
Hi Thinh,

On 7/12/2022 6:42 PM, Thinh Nguyen wrote:
On 7/12/2022, Wesley Cheng wrote:
Hi Thinh,

On 7/8/2022 6:58 PM, Thinh Nguyen wrote:
On 7/8/2022, Wesley Cheng wrote:
If soft disconnect is in progress, allow the endxfer command to be
without this, there is an issue where the stop active transfer call
(during pullup disable) wouldn't actually issue the endxfer command,
while clearing the DEP flag.

In addition, if the DWC3_EP_DELAY_STOP flag was set before soft
started (i.e. from the dequeue path), ensure that when the EP0
completes during soft disconnect, to issue the endxfer with the force
parameter set, as it does not expect a command complete event.

Signed-off-by: Wesley Cheng <quic_wcheng@xxxxxxxxxxx>
    drivers/usb/dwc3/ep0.c    | 3 +--
    drivers/usb/dwc3/gadget.c | 5 ++++-
    2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/drivers/usb/dwc3/ep0.c b/drivers/usb/dwc3/ep0.c
index 506ef717fdc0..5851b0e9db0a 100644
--- a/drivers/usb/dwc3/ep0.c
+++ b/drivers/usb/dwc3/ep0.c
@@ -290,8 +290,7 @@ void dwc3_ep0_out_start(struct dwc3 *dwc)
            if (!(dwc3_ep->flags & DWC3_EP_DELAY_STOP))
    -        dwc3_ep->flags &= ~DWC3_EP_DELAY_STOP;
-        dwc3_stop_active_transfer(dwc3_ep, true, true);
+        dwc3_stop_active_transfer(dwc3_ep, true, dwc->connected);
    diff --git a/drivers/usb/dwc3/gadget.c
index bd40608b19df..fba2797ad9ae 100644
--- a/drivers/usb/dwc3/gadget.c
+++ b/drivers/usb/dwc3/gadget.c
@@ -3696,8 +3696,10 @@ void dwc3_stop_active_transfer(struct dwc3_ep
*dep, bool force,
        if (dep->number <= 1 && dwc->ep0state != EP0_DATA_PHASE)
    +    if (interrupt && (dep->flags & DWC3_EP_DELAY_STOP))
+        return;
        if (!(dep->flags & DWC3_EP_TRANSFER_STARTED) ||
-        (dep->flags & DWC3_EP_DELAY_STOP) ||
            (dep->flags & DWC3_EP_END_TRANSFER_PENDING))
    @@ -3744,6 +3746,7 @@ void dwc3_stop_active_transfer(struct
dwc3_ep *dep, bool force,
        __dwc3_stop_active_transfer(dep, force, interrupt);
    +    dep->flags &= ~DWC3_EP_DELAY_STOP;

Can we clear this flag in __dwc3_stop_active_transfer(). It should
if End Transfer command was sent.

I wanted to make sure that we weren't modifying the DEP flags outside
of a spin lock.  Patch#3 modifies it where we unlock before calling
__dwc3_stop_active_transfer(), so we can allow the dwc3 threaded IRQ
handle events while the cmd status polling happens.

Maybe we can unlock/lock the dwc3->lock inside
__dwc3_stop_active_transfer() and that way we can ensure DEP flags are
modified properly?

I didn't realize that you unlock/lock when calling
__dwc3_stop_active_transfer(). We'd need to be careful if we want to
unlock/lock it, and avoid it all together if possible. It can be easily
overlooked just like dwc3_gadget_giveback().

What issue did you see without doing this?

I saw endxfer timeout issues if I didn't do it.  If we keep the lock
held, then the DWC3 event processing would be blocked across the
entire time we are waiting for the command act to clear.  With
unlocking before polling, then at least we're still able to handle the
EP0 events that are pending.

It was definitely one of the harder scenarios to reproduce.  The main
patch series which resolved a lot of the issues early on was patch#1.
After adding that the other issues are seen maybe after a day or so of

I see. The soft-disconnect still go through right?
Can you try this and see if this works for you?

diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
index 13e4f1a03417..4f67a484c490 100644
--- a/drivers/usb/dwc3/gadget.c
+++ b/drivers/usb/dwc3/gadget.c
@@ -451,7 +451,9 @@ int dwc3_send_gadget_ep_cmd(struct dwc3_ep *dep,
unsigned int cmd,

        dwc3_writel(dep->regs, DWC3_DEPCMD, cmd);

-       if (!(cmd & DWC3_DEPCMD_CMDACT)) {
+       if (!(cmd & DWC3_DEPCMD_CMDACT) ||
+            !(cmd & DWC3_DEPCMD_CMDIOC))) {
                ret = 0;
                goto skip_status;

We can set it and forget it when we're about to de-initialize the
controller. Just like what we've already done by not waiting for the End
Transfer endpoint command completion interrupt.

I can see this working for the soft disconnect case, but we'll need to address it differently for the EP dequeue scenario where IOC/interrupt is required to clean up the TRBs. This is one of the reasons why I went with the spinlock approach, since it would apply to both situations.

Wesley Cheng