On 17 November 2017 17:20, Guenter Roeck wrote:
> On Thu, Nov 16, 2017 at 04:28:11PM +0000, Adam Thomson wrote:
> > The expectation in the FUSB302 driver is that a TX_SUCCESS event
> > should occur after a message has been sent, but before a GCRCSENT
> > event is raised to indicate successful receipt of a message from
> > the partner. However in some circumstances it is possible to see
> > the hardware raise a GCRCSENT event before a TX_SUCCESS event
> > is raised. The upshot of this is that the GCRCSENT handling portion
> > of code ends up reporting the GoodCRC message to TCPM because the
> > TX_SUCCESS event hasn't yet arrived to trigger a consumption of it.
> > When TX_SUCCESS is then raised by the chip it ends up consuming the
> > actual message that was meant for TCPM, and this incorrect sequence
> > results in a hard reset from TCPM.
> >
> > To avoid this problem, this commit moves all FIFO reading to be
> > done based on a GCRCSENT event, and when reading from the FIFO
> > any GoodCRC messages read in are discarded so only valid messages
> > are reported to TCPM.
> >
> > Changes in v2:
> > - Remove erroneous extended header check
> >
> > Patch is based on Linux next-20171114 to include move out of staging.
> >
> > Signed-off-by: Adam Thomson <adam.thomson.opensou...@diasemi.com>
> > ---
> > drivers/usb/typec/fusb302/fusb302.c | 16 ++++++++++------
> > 1 file changed, 10 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/usb/typec/fusb302/fusb302.c
> b/drivers/usb/typec/fusb302/fusb302.c
> > index 72cb060..ddf88f0 100644
> > --- a/drivers/usb/typec/fusb302/fusb302.c
> > +++ b/drivers/usb/typec/fusb302/fusb302.c
> > @@ -1650,12 +1650,6 @@ static irqreturn_t fusb302_irq_intn(int irq, void
> *dev_id)
> >
> > if (interrupta & FUSB_REG_INTERRUPTA_TX_SUCCESS) {
> > fusb302_log(chip, "IRQ: PD tx success");
> > - /* read out the received good CRC */
> > - ret = fusb302_pd_read_message(chip, &pd_msg);
> > - if (ret < 0) {
> > - fusb302_log(chip, "cannot read in GCRC, ret=%d", ret);
> > - goto done;
> > - }
>
> If multiple "Good CRC" messages are received in a row, they won't be read from
> the chip, which might result in a buffer overflow.
Thanks for your comments. I had considered this since sending the patch. In
reality I think we'd need to send 11 consecutive messages to which we receive
11 GoodCRCs without receiving a real response message from the partner to clear
the FIFO. Having looked again at the PD spec that seemed really unlikely, but...
> It might be better to always read all pending messages and handle it depending
> on the message type. Something along the line of
>
> while (interrupts & (FUSB_REG_INTERRUPTA_TX_SUCCESS |
> FUSB_REG_INTERRUPTB_GCRCSENT)) {
> ret = fusb302_pd_read_message(chip, &pd_msg);
> if (ret < 0)
> goto done;
> if (msg_type == good CRC) {
> tcpm_pd_transmit_complete(chip->tcpm_port,
> TCPC_TX_SUCCESS);
> interrupts &= ~FUSB_REG_INTERRUPTA_TX_SUCCESS;
> } else {
> tcpm_pd_receive(chip->tcpm_port, &pd_msg);
> interrupts &= ~FUSB_REG_INTERRUPTB_GCRCSENT;
> }
> }
>
> Guenter
...I do prefer your approach as it doesn't leave anything to chance. I'll
respin accordingly.
>
> > tcpm_pd_transmit_complete(chip->tcpm_port, TCPC_TX_SUCCESS);
> > }
> >
> > @@ -1671,12 +1665,22 @@ static irqreturn_t fusb302_irq_intn(int irq, void
> *dev_id)
> >
> > if (interruptb & FUSB_REG_INTERRUPTB_GCRCSENT) {
> > fusb302_log(chip, "IRQ: PD sent good CRC");
> > +retry:
> > ret = fusb302_pd_read_message(chip, &pd_msg);
> > if (ret < 0) {
> > fusb302_log(chip,
> > "cannot read in PD message, ret=%d", ret);
> > goto done;
> > }
> > +
> > + /*
> > + * Check to make sure we've not read off a GoodCRC message.
> > + * If so then read again to retrieve expected message
> > + */
> > + if ((!pd_header_cnt_le(pd_msg.header)) &&
> > + (pd_header_type_le(pd_msg.header) == PD_CTRL_GOOD_CRC))
> > + goto retry;
> > +
> > tcpm_pd_receive(chip->tcpm_port, &pd_msg);
> > }
> > done:
> > --
> > 1.9.1
> >
