Re: [PATCH 1/2] mmc: cqhci: Add a quirk to clear stale TC

[Adrian Hunter]

On 1/11/23 13:31, Kornel Dulęba wrote:
> On Mon, Oct 30, 2023 at 8:31 PM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>>
>> On 27/10/23 17:56, Kornel Dulęba wrote:
>>> This fix addresses a stale task completion event issued right after the
>>> CQE recovery. As it's a hardware issue the fix is done in form of a
>>> quirk.
>>>
>>> When error interrupt is received the driver runs recovery logic is run.
>>> It halts the controller, clears all pending tasks, and then re-enables
>>> it. On some platforms a stale task completion event is observed,
>>> regardless of the CQHCI_CLEAR_ALL_TASKS bit being set.
>>>
>>> This results in either:
>>> a) Spurious TC completion event for an empty slot.
>>> b) Corrupted data being passed up the stack, as a result of premature
>>>    completion for a newly added task.
>>>
>>> To fix that re-enable the controller, clear task completion bits,
>>> interrupt status register and halt it again.
>>> This is done at the end of the recovery process, right before interrupts
>>> are re-enabled.
>>>
>>> Signed-off-by: Kornel Dulęba <korneld@xxxxxxxxxxxx>
>>> ---
>>>  drivers/mmc/host/cqhci-core.c | 42 +++++++++++++++++++++++++++++++++++
>>>  drivers/mmc/host/cqhci.h      |  1 +
>>>  2 files changed, 43 insertions(+)
>>>
>>> diff --git a/drivers/mmc/host/cqhci-core.c b/drivers/mmc/host/cqhci-core.c
>>> index b3d7d6d8d654..e534222df90c 100644
>>> --- a/drivers/mmc/host/cqhci-core.c
>>> +++ b/drivers/mmc/host/cqhci-core.c
>>> @@ -1062,6 +1062,45 @@ static void cqhci_recover_mrqs(struct cqhci_host *cq_host)
>>>  /* CQHCI could be expected to clear it's internal state pretty quickly */
>>>  #define CQHCI_CLEAR_TIMEOUT          20
>>>
>>> +/*
>>> + * During CQE recovery all pending tasks are cleared from the
>>> + * controller and its state is being reset.
>>> + * On some platforms the controller sets a task completion bit for
>>> + * a stale(previously cleared) task right after being re-enabled.
>>> + * This results in a spurious interrupt at best and corrupted data
>>> + * being passed up the stack at worst. The latter happens when
>>> + * the driver enqueues a new request on the problematic task slot
>>> + * before the "spurious" task completion interrupt is handled.
>>> + * To fix it:
>>> + * 1. Re-enable controller by clearing the halt flag.
>>> + * 2. Clear interrupt status and the task completion register.
>>> + * 3. Halt the controller again to be consistent with quirkless logic.
>>> + *
>>> + * This assumes that there are no pending requests on the queue.
>>> + */
>>> +static void cqhci_quirk_clear_stale_tc(struct cqhci_host *cq_host)
>>> +{
>>> +     u32 reg;
>>> +
>>> +     WARN_ON(cq_host->qcnt);
>>> +     cqhci_writel(cq_host, 0, CQHCI_CTL);
>>> +     if ((cqhci_readl(cq_host, CQHCI_CTL) & CQHCI_HALT)) {
>>> +             pr_err("%s: cqhci: CQE failed to exit halt state\n",
>>> +                     mmc_hostname(cq_host->mmc));
>>> +     }
>>> +     reg = cqhci_readl(cq_host, CQHCI_TCN);
>>> +     cqhci_writel(cq_host, reg, CQHCI_TCN);
>>> +     reg = cqhci_readl(cq_host, CQHCI_IS);
>>> +     cqhci_writel(cq_host, reg, CQHCI_IS);
>>> +
>>> +     /*
>>> +      * Halt the controller again.
>>> +      * This is only needed so that we're consistent across quirk
>>> +      * and quirkless logic.
>>> +      */
>>> +     cqhci_halt(cq_host->mmc, CQHCI_FINISH_HALT_TIMEOUT);
>>> +}
>>
>> Thanks a lot for tracking this down!
>>
>> It could be that the "un-halt" starts a task, so it would be
>> better to force the "clear" to work if possible, which
>> should be the case if CQE is disabled.
>>
>> Would you mind trying the code below?  Note the increased
>> CQHCI_START_HALT_TIMEOUT helps avoid trying to clear tasks
>> when CQE has not halted.
> 
> Sure, I'll try it out tomorrow, as I don't have access to the DUT today.
> BTW do we even need to halt the controller in the recovery_finish logic?
> It has already been halted in recovery_start, I guess it could be
> there in case the recovery_start halt didn't work.
> But in that case shouldn't we do this disable/re-enable dance in recovery_start?

"Halt" might be waiting on an operation to finish, so the STOP
command is meant to help bring that to a conclusion.