Re: [PATCH] usb: cdns3: remove fetched trb from cache before dequeuing

[Peter Chen <peter.chen@kernel.org>]

On 23-01-09 05:38:31, Pawel Laszczak wrote:
> >
> >On Thu, Nov 17, 2022 at 8:27 PM Pawel Laszczak <pawell@cadence.com>
> >wrote:
> >>
> >> >
> >> >On Tue, Nov 15, 2022 at 6:01 PM Pawel Laszczak <pawell@cadence.com>
> >> >wrote:
> >> >>
> >> >> After doorbell DMA fetches the TRB. If during dequeuing request
> >> >> driver changes NORMAL TRB to LINK TRB but doesn't delete it from
> >> >> controller cache then controller will handle cached TRB and packet can be
> >lost.
> >> >>
> >> >> The example scenario for this issue looks like:
> >> >> 1. queue request - set doorbell
> >> >> 2. dequeue request
> >> >> 3. send OUT data packet from host
> >> >> 4. Device will accept this packet which is unexpected 5. queue new
> >> >> request - set doorbell 6. Device lost the expected packet.
> >> >>
> >> >> By setting DFLUSH controller clears DRDY bit and stop DMA transfer.
> >> >>
> >> >> Fixes: 7733f6c32e36 ("usb: cdns3: Add Cadence USB3 DRD Driver")
> >> >> cc: <stable@vger.kernel.org>
> >> >> Signed-off-by: Pawel Laszczak <pawell@cadence.com>
> >> >> ---
> >> >>  drivers/usb/cdns3/cdns3-gadget.c | 12 ++++++++++++
> >> >>  1 file changed, 12 insertions(+)
> >> >>
> >> >> diff --git a/drivers/usb/cdns3/cdns3-gadget.c
> >> >> b/drivers/usb/cdns3/cdns3-gadget.c
> >> >> index 5adcb349718c..ccfaebca6faa 100644
> >> >> --- a/drivers/usb/cdns3/cdns3-gadget.c
> >> >> +++ b/drivers/usb/cdns3/cdns3-gadget.c
> >> >> @@ -2614,6 +2614,7 @@ int cdns3_gadget_ep_dequeue(struct usb_ep
> >*ep,
> >> >>         u8 req_on_hw_ring = 0;
> >> >>         unsigned long flags;
> >> >>         int ret = 0;
> >> >> +       int val;
> >> >>
> >> >>         if (!ep || !request || !ep->desc)
> >> >>                 return -EINVAL;
> >> >> @@ -2649,6 +2650,13 @@ int cdns3_gadget_ep_dequeue(struct usb_ep
> >> >*ep,
> >> >>
> >> >>         /* Update ring only if removed request is on pending_req_list list */
> >> >>         if (req_on_hw_ring && link_trb) {
> >> >> +               /* Stop DMA */
> >> >> +               writel(EP_CMD_DFLUSH, &priv_dev->regs->ep_cmd);
> >> >> +
> >> >> +               /* wait for DFLUSH cleared */
> >> >> +               readl_poll_timeout_atomic(&priv_dev->regs->ep_cmd, val,
> >> >> +                                         !(val & EP_CMD_DFLUSH),
> >> >> + 1, 1000);
> >> >> +
> >> >>                 link_trb->buffer = cpu_to_le32(TRB_BUFFER(priv_ep-
> >> >>trb_pool_dma +
> >> >>                         ((priv_req->end_trb + 1) * TRB_SIZE)));
> >> >>                 link_trb->control =
> >> >> cpu_to_le32((le32_to_cpu(link_trb->control) & TRB_CYCLE) | @@
> >> >>-2660,6
> >> >> +2668,10 @@ int cdns3_gadget_ep_dequeue(struct usb_ep *ep,
> >> >>
> >> >>         cdns3_gadget_giveback(priv_ep, priv_req, -ECONNRESET);
> >> >>
> >> >> +       req = cdns3_next_request(&priv_ep->pending_req_list);
> >> >> +       if (req)
> >> >> +               cdns3_rearm_transfer(priv_ep, 1);
> >> >> +
> >> >
> >> >Why the above changes are needed?
> >> >
> >>
> >> Do you mean the last line or this patch?
> >>
> >> Last line:
> >> DMA is stopped, so driver arm the queued transfers
> >>
> >
> >Sorry, I have been very busy recently, so the response may not be in time.
> >I mean why it needs to re-arm the transfers after DMA is stopped?
> 
> Because driver can have queued more transfers. Only one of them are
> dequeued. In the vast majority of the rest request will be removed in the
> next steps, but there can be case in which we have queued e.g. 10 usb requests
> and only one of them will be removed. In such case the driver can stuck.
> To avoid this driver, rearm the endpoint if there are other transfer
> in transfer ring.
> 

Since this logic (re-arm the pending request) is different with current
one, please test it well to avoid other use cases. After you have fully
tested, feel free to add my ack:

Acked-by: Peter Chen <peter.chen@kernel.org>

Peter

> Regards,
> Pawel
> 
> >
> >
> >> If you means this patch:
> >> Issue was detected by customer test. I don’t know whether it was only
> >> test or the real application.
> >>
> >> The problem happens because user application queued the transfer
> >> (endpoint has been armed), so controller fetch the TRB.
> >> When user application removed this request the TRB was still processed
> >> by controller. If at that time the host will send data packet then
> >> controller will accept it, but it shouldn't because the usb_request
> >> associated with TRB cached by controller was removed.
> >> To force the controller to drop this TRB DFLUSH is required.
> >>
> >> Pawel
> >>
> >> >
> >> >>  not_found:
> >> >>         spin_unlock_irqrestore(&priv_dev->lock, flags);
> >> >>         return ret;
> >> >> --
> >> >> 2.25.1
> >> >>

-- 

Thanks,
Peter Chen