Re: [PATCH v8 6/6] serial: 8250: retain DMA resources when termination fails

[Andy Shevchenko]

On Fri, Oct 30, 2026 at 12:08:52PM -0600, James Hilliard wrote:
> On Fri, Oct 30, 2026 at 11:54 AM Andy Shevchenko
> <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > On Fri, Oct 30, 2026 at 11:09:53AM -0600, James Hilliard wrote:
> > > On Fri, Oct 30, 2026 at 3:40 AM Andy Shevchenko
> > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > > > On Thu, Oct 29, 2026 at 11:19:59AM -0600, James Hilliard wrote:
> > > > > On Thu, Oct 29, 2026 at 8:24 AM Andy Shevchenko
> > > > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > > > > > On Thu, Oct 29, 2026 at 01:37:02AM -0600, James Hilliard wrote:

...

> > > > > > > +     while (rx_ret || tx_ret) {
> > > > > > > +             dev_err_ratelimited(p->port.dev,
> > > > > > > +                                 "DMA termination failed: RX %d, TX %d; retrying\n",
> > > > > > > +                                 rx_ret, tx_ret);
> > > > > >
> > > > > > > +             msleep(100);
> > > > > >
> > > > > > Huh?!
> > > > > >
> > > > > > > +             if (rx_ret)
> > > > > > > +                     rx_ret = dmaengine_terminate_sync(dma->rxchan);
> > > > > > > +             if (tx_ret)
> > > > > > > +                     tx_ret = dmaengine_terminate_sync(dma->txchan);
> > > > > > > +     }
> > > > > >
> > > > > > This is unacceptable. Sleeping is okay, but value should really match the
> > > > > > baudspeed and other parameters of the transfer.
> > > > >
> > > > > This sleep is backoff after dmaengine_terminate_sync() has returned
> > > > > an error, rather than an estimate of UART transmission time. Each
> > > > > retry asks the provider to stop the channel again; elapsed time is
> > > > > never treated as evidence that the resources are safe to free.
> > > > >
> > > > > The particular 100 ms interval has no hardware-derived justification.
> > > > > However, I don't yet see a sound way to derive the retry cadence from
> > > > > baud rate either. RX can be waiting for input, TX can be held by flow
> > > > > control, and callback retirement is a separate requirement from
> > > > > moving bytes through the UART.
> > > > >
> > > > > Do you have a particular provider failure mode in mind where retrying
> > > > > termination should follow the serial transfer timing? I'd like to
> > > > > base the interval on the condition that can make termination succeed,
> > > > > rather than substitute a wire-time estimate for an arbitrary delay.
> > > >
> > > > Termination of DMA sometimes require to clean up internal FIFOs and other state
> > > > machines. That's what happens on HW side. In the software waiting for too much
> > > > may lead to the very big undesired latencies. The value of 100ms sounds insane
> > > > to me for the HSUART cases (megabaud range of speed). Also note, not every PCB
> > > > design uses HW flow control, so this may be even worse for them.
> > >
> > > The 100 ms interval can add unnecessary teardown latency if a failed
> > > termination becomes retryable much sooner. There is no hardware-derived
> > > justification for that particular value.
> > >
> > > The FIFO/state-machine draining is part of the provider's stop sequence.
> > > For example, DW performs its FIFO-empty polling inside
> > > dwc_terminate_all(). If termination succeeds, this patch proceeds without
> > > the retry sleep, regardless of baud rate. The sleep is only reached
> > > after an error during DMA release, not during normal RX delivery.
> > >
> > > Baud rate can influence how quickly a transmit FIFO drains, even without
> > > flow control. However, the termination error does not tell 8250 whether
> > > that drain is still progressing, how much remains, or whether the
> > > controller needs a different recovery operation. That distinction matters
> > > when choosing a retry interval: a shorter interval reduces the added
> > > latency, but does not establish that retrying can recover the channel.
> > >
> > > There is also a specific API detail here: dmaengine_terminate_sync()
> > > returns immediately if device_terminate_all() fails, without reaching
> > > dmaengine_synchronize(). Consequently, an error leaves callback draining
> > > unconfirmed as well. The retry cadence can change, but resource release
> > > still requires successful termination and synchronization.
> >
> > Most likely this is AI slop. I can't see the justification for 100ms.
> > Can you explain that?
> 
> 100 ms was an arbitrary backoff to avoid a tight retry loop if
> termination keeps failing. It was not derived from the baud rate,
> FIFO depth, or a DMA provider's recovery timing.
> 
> That motivation explains sleeping between retries, but not choosing
> 100 ms. I don't have evidence supporting that particular interval.

Yep. And that's what my point is about. Any big sleeps has to be explained
and justified.

> > > > > Regardless of the retry interval, shutdown cannot release the buffers
> > > > > until the provider establishes that DMA and its callbacks have stopped.

-- 
With Best Regards,
Andy Shevchenko