Re: [PATCH v4 2/4] platform/x86: intel_scu_ipc: Check status upon timeout in ipc_wait_for_interrupt()

["Ilpo Järvinen" <ilpo.jarvinen@linux.intel.com>]

[-- Attachment #1: Type: text/plain, Size: 3040 bytes --]

On Mon, 18 Sep 2023, Hans de Goede wrote:
> On 9/15/23 15:49, Ilpo Järvinen wrote:
> > On Wed, 13 Sep 2023, Stephen Boyd wrote:
> > 
> >> It's possible for the completion in ipc_wait_for_interrupt() to timeout,
> >> simply because the interrupt was delayed in being processed. A timeout
> >> in itself is not an error. This driver should check the status register
> >> upon a timeout to ensure that scheduling or interrupt processing delays
> >> don't affect the outcome of the IPC return value.
> >>
> >>  CPU0                                                   SCU
> >>  ----                                                   ---
> >>  ipc_wait_for_interrupt()
> >>   wait_for_completion_timeout(&scu->cmd_complete)
> >>   [TIMEOUT]                                             status[IPC_STATUS_BUSY]=0
> >>
> >> Fix this problem by reading the status bit in all cases, regardless of
> >> the timeout. If the completion times out, we'll assume the problem was
> >> that the IPC_STATUS_BUSY bit was still set, but if the status bit is
> >> cleared in the meantime we know that we hit some scheduling delay and we
> >> should just check the error bit.
> > 
> > Hi,
> > 
> > I don't understand the intent here. What prevents IPC_STATUS_BUSY from 
> > changing right after you've read it in ipc_read_status(scu)? Doesn't that 
> > end you exactly into the same situation where the returned value is stale 
> > so I cannot see how this fixes anything, at best it just plays around the 
> > race window that seems to still be there after this fix?
> 
> As I understand it the problem before was that the function would
> return -ETIMEDOUT; purely based on wait_for_completion_timeout()
> without ever actually checking the BUSY bit:
> 
> Old code:
> 
> 	if (!wait_for_completion_timeout(&scu->cmd_complete, IPC_TIMEOUT))
> 		return -ETIMEDOUT;
> 
> This allows for a scenario where when the IRQ processing got delayed
> (on say another core) causing the timeout to trigger,
> ipc_wait_for_interrupt() would return -ETIMEDOUT even though
> the BUSY flag was already cleared by the SCU.
> 
> This patch adds an explicit check for the BUSY flag after
> the wait_for_completion(), rather then relying on the
> wait_for_completion() return value which implies things
> are still busy.

Oh, I see, it's because the code is waiting for the completion rather than
the actual condition.

> As for "What prevents IPC_STATUS_BUSY from 
> changing right after you've read it in ipc_read_status(scu)?"
> 
> AFAICT in this code path the bit is only ever supposed to go
> from being set (busy) to unset (not busy), not the other
> way around since no new commands can be submitted until
> this function has completed. So that scenario cannot happen.

This is not what I meant.

I meant that if the code has decided to return -ETIMEDOUT, the status bit 
still change at that point which makes the return value to not match. This 
race is still there and given the changelog was a bit sparse on what race 
it was fixing I ended up noticing this detail.


-- 
 i.