Re: [PATCH v2] tpm,tpm_tis: Handle interrupt storm

[Jarkko Sakkinen]

On Fri Jun 9, 2023 at 7:03 PM EEST, Lino Sanfilippo wrote:
>
> Hi,
>
> On 09.06.23 16:33, Jarkko Sakkinen wrote:
>
> > 
> > On Tue May 30, 2023 at 8:47 PM EEST, Lino Sanfilippo wrote:
> >> From: Lino Sanfilippo <l.sanfilippo@xxxxxxxxxx>
> >>
> >> After activation of interrupts for TPM TIS drivers 0-day reports an
> >> interrupt storm on an Inspur NF5180M6/NF5180M6 server.
> >>
> >> Fix this by detecting the storm and falling back to polling:
> >> Count the number of unhandled interrupts within a 10 ms time interval. In
> >> case that more than 1000 were unhandled deactivate interrupts entirely,
> >> deregister the handler and use polling instead.
> >>
> >> The storm detection logic equals the implementation in note_interrupt()
> >> which uses timestamps and counters stored in struct irq_desc. Since this
> >> structure is private to the generic interrupt core the TPM TIS core uses
> >> its own timestamps and counters. Furthermore the TPM interrupt handler
> >> always returns IRQ_HANDLED to prevent the generic interrupt core from
> >> processing the interrupt storm.
> >>
> >> Since the interrupt deregistration function devm_free_irq() waits for all
> >> interrupt handlers to finish, only trigger a worker in the interrupt
> >> handler and do the unregistration in the worker to avoid a deadlock.
> >>
> >> Reported-by: kernel test robot <yujie.liu@xxxxxxxxx>
> >> Closes: https://lore.kernel.org/oe-lkp/202305041325.ae8b0c43-yujie.liu@xxxxxxxxx/
> >> Suggested-by: Lukas Wunner <lukas@xxxxxxxxx>
> >> Signed-off-by: Lino Sanfilippo <l.sanfilippo@xxxxxxxxxx>
> >> ---
> > 
> > Sorry for the latency. I've moved home office to a new location,
> > which has caused ~2 week lag. Unfortunate timing.
> > 
>
>
> No prob :)
>
>
> >>  drivers/char/tpm/tpm_tis_core.c | 93 ++++++++++++++++++++++++++++-----
> >>  drivers/char/tpm/tpm_tis_core.h |  4 ++
> >>  2 files changed, 85 insertions(+), 12 deletions(-)
> >>
> >> diff --git a/drivers/char/tpm/tpm_tis_core.c b/drivers/char/tpm/tpm_tis_core.c
> >> index 558144fa707a..7ae8228e803f 100644
> >> --- a/drivers/char/tpm/tpm_tis_core.c
> >> +++ b/drivers/char/tpm/tpm_tis_core.c
> >> @@ -468,25 +468,32 @@ static int tpm_tis_send_data(struct tpm_chip *chip, const u8 *buf, size_t len)
> >>       return rc;
> >>  }
> >>
> >> +static void __tpm_tis_disable_interrupts(struct tpm_chip *chip)
> >> +{
> >> +     struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
> >> +     u32 intmask = 0;
> >> +
> >> +     tpm_tis_read32(priv, TPM_INT_ENABLE(priv->locality), &intmask);
> >> +     intmask &= ~TPM_GLOBAL_INT_ENABLE;
> >> +
> >> +     tpm_tis_request_locality(chip, 0);
> >> +     tpm_tis_write32(priv, TPM_INT_ENABLE(priv->locality), intmask);
> >> +     tpm_tis_relinquish_locality(chip, 0);
> >> +
> >> +     chip->flags &= ~TPM_CHIP_FLAG_IRQ;
> >> +}
> >> +
> >>  static void disable_interrupts(struct tpm_chip *chip)
> > 
> > Add tpm_ prefix here too. It makes tracing/grepping/etc so much nicer.
>
> Ok.
>
> > 
> >>  {
> >>       struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
> >> -     u32 intmask;
> > 
> > int_mask is more readable
>
> Ok.
>
> > 
> >> -     int rc;
> >>
> >>       if (priv->irq == 0)
> >>               return;
> >>
> >> -     rc = tpm_tis_read32(priv, TPM_INT_ENABLE(priv->locality), &intmask);
> >> -     if (rc < 0)
> >> -             intmask = 0;
> >> -
> >> -     intmask &= ~TPM_GLOBAL_INT_ENABLE;
> >> -     rc = tpm_tis_write32(priv, TPM_INT_ENABLE(priv->locality), intmask);
> >> +     __tpm_tis_disable_interrupts(chip);
> >>
> >>       devm_free_irq(chip->dev.parent, priv->irq, chip);
> >>       priv->irq = 0;
> >> -     chip->flags &= ~TPM_CHIP_FLAG_IRQ;
> >>  }
> >>
> >>  /*
> >> @@ -752,6 +759,53 @@ static bool tpm_tis_req_canceled(struct tpm_chip *chip, u8 status)
> >>       return status == TPM_STS_COMMAND_READY;
> >>  }
> >>
> >> +static void tpm_tis_reenable_polling(struct tpm_chip *chip)
> >> +{
> >> +     struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
> >> +
> >> +     dev_warn(&chip->dev, FW_BUG
> >> +              "TPM interrupt storm detected, polling instead\n");
> >> +
> >> +     __tpm_tis_disable_interrupts(chip);
> >> +
> >> +     /*
> >> +      * devm_free_irq() must not be called from within the interrupt handler,
> >> +      * since this function waits for running handlers to finish and thus it
> >> +      * would deadlock. Instead trigger a worker that takes care of the
> >> +      * unregistration.
> >> +      */
> >> +     schedule_work(&priv->free_irq_work);
> >> +}
> >> +
> >> +static irqreturn_t tpm_tis_check_for_interrupt_storm(struct tpm_chip *chip)
> >> +{
> >> +     struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
> >> +     const unsigned int MAX_UNHANDLED_IRQS = 1000;
> > 
> > Please declare this in the beginning of file because it is non-empirical
> > tuning parameter. I do not want it to be buried here. It is now as good
> > as a magic number.
> > 
> > Or perhaps even tpm_tis_core.h?
> > 
>
> For now that constant is only used in tpm_tis_core.c. So I would favor to define it there.
>
> > Why MAX_UNHANDLED_IRQS is exactly 1000 and not 1? I would rollback eagerly.
>
>
> Because the IRQ line may be shared with another device which has raised the
> interrupt instead of the TPM. So unhandled interrupts may be legit.

I understand that being exact here is impossible. So let's stick to this
but please move the constant to the tpm_tis_core.c with the TPM_ prefix
because it is an essential tuning parameter.

BR, Jarkko