Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Mime-Version: 1.0
Content-Transfer-Encoding: quoted-printable
Content-Type: text/plain; charset=UTF-8
Date:   Wed, 24 May 2023 06:59:59 +0300
Message-Id: <CSU7HHIKIE8Q.17FAFU7I0XHSV@suppilovahvero>
Cc:     <jsnitsel@redhat.com>, <hdegoede@redhat.com>,
        <oe-lkp@lists.linux.dev>, <lkp@intel.com>,
        <peter.ujfalusi@linux.intel.com>, <peterz@infradead.org>,
        <linux@mniewoehner.de>, <linux-integrity@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>, <l.sanfilippo@kunbus.com>,
        <lukas@wunner.de>, <p.rosenberger@kunbus.com>
Subject: Re: [PATCH 1/2] tpm, tpm_tis: Handle interrupt storm
From:   "Jarkko Sakkinen" <jarkko@kernel.org>
To:     "Jarkko Sakkinen" <jarkko@kernel.org>,
        "Lino Sanfilippo" <LinoSanfilippo@gmx.de>, <peterhuewe@gmx.de>,
        <jgg@ziepe.ca>
References: <20230522143105.8617-1-LinoSanfilippo@gmx.de>
 <CSU7G2UZSZ8K.22EGXU5CJTZBB@suppilovahvero>
In-Reply-To: <CSU7G2UZSZ8K.22EGXU5CJTZBB@suppilovahvero>
Precedence: bulk

On Wed May 24, 2023 at 6:58 AM EEST, Jarkko Sakkinen wrote:
> Hi,
>
> Sorry, some minor glitches.
>
> On Mon May 22, 2023 at 5:31 PM EEST, Lino Sanfilippo wrote:
> > From: Lino Sanfilippo <l.sanfilippo@kunbus.com>
> >
> > Commit e644b2f498d2 ("tpm, tpm_tis: Enable interrupt test") enabled
> > interrupts instead of polling on all capable TPMs. Unfortunately, on so=
me
> > products the interrupt line is either never asserted or never deasserte=
d.
>
> Use Reported-by and Closes tag and remove this paragraph.
>
> In Closes link instead from lore the email where the bug was reported.
>
> > The former causes interrupt timeouts and is detected by
> > tpm_tis_core_init(). The latter results in interrupt storms.
>
> Please describe instead the system where the bug was realized. Don't
> worry about speculative descriptions. We only deal with ones actually
> realized.
>
> > Recent reports concern the Lenovo ThinkStation P360 Tiny, Lenovo ThinkP=
ad
> > L490 and Inspur NF5180M6:
> >
> > https://lore.kernel.org/linux-integrity/20230511005403.24689-1-jsnitsel=
@redhat.com/
> > https://lore.kernel.org/linux-integrity/d80b180a569a9f068d3a2614f062cfa=
3a78af5a6.camel@kernel.org/
>
> Please remove all of this, as the fixes have been handled. Let's keep
> the commit message focused.
>
> > The current approach to avoid those storms is to disable interrupts by
> > adding a DMI quirk for the concerned device.
> >
> > However this is a maintenance burden in the long run, so use a generic
> > approach:
> >
> > Detect an interrupt storm by counting the number of unhandled interrupt=
s
> > within a 10 ms time interval. In case that more than 1000 were unhandle=
d
> > deactivate interrupts, deregister the handler and fall back to polling.
> >
> > This equals the implementation that handles interrupt storms in
> > note_interrupt() by means of timestamps and counters in struct irq_desc=
.
> > However the function to access this structure is private so the logic h=
as
> > to be reimplemented in the TPM TIS core.
>
> I only now found out that this is based on kernel/irq/spurious.c code
> partly. Why this was unmentioned?
>
> That would make this already more legitimate because it is based
> on field tested metrics.
>
> Then we only have to discuss about counter.
>
> > routine trigger a worker thread that executes the unregistration.
> >
> > Suggested-by: Lukas Wunner <lukas@wunner.de>
> > Signed-off-by: Lino Sanfilippo <l.sanfilippo@kunbus.com>
> > ---
> >  drivers/char/tpm/tpm_tis_core.c | 71 +++++++++++++++++++++++++++++++--
> >  drivers/char/tpm/tpm_tis_core.h |  6 +++
> >  2 files changed, 74 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/char/tpm/tpm_tis_core.c b/drivers/char/tpm/tpm_tis=
_core.c
> > index 558144fa707a..458ebf8c2f16 100644
> > --- a/drivers/char/tpm/tpm_tis_core.c
> > +++ b/drivers/char/tpm/tpm_tis_core.c
> > @@ -752,6 +752,55 @@ static bool tpm_tis_req_canceled(struct tpm_chip *=
chip, u8 status)
> >  	return status =3D=3D TPM_STS_COMMAND_READY;
> >  }
> > =20
> > +static void tpm_tis_handle_irq_storm(struct tpm_chip *chip)
> > +{
> > +	struct tpm_tis_data *priv =3D dev_get_drvdata(&chip->dev);
> > +	int intmask =3D 0;
> > +
> > +	dev_err(&chip->dev, HW_ERR
> > +		"TPM interrupt storm detected, polling instead\n");
>
> Degrading this to warn is fine because it is legit behaviour in a
> sense.
>
> > +
> > +	tpm_tis_read32(priv, TPM_INT_ENABLE(priv->locality), &intmask);
> > +
> > +	intmask &=3D ~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 &=3D ~TPM_CHIP_FLAG_IRQ;
> > +
> > +	/*
> > +	 * We must not call devm_free_irq() from within the interrupt handler=
,
>
> Never use "we" form. Always use either:
>
> 1. Imperative
> 2. Passive
>
> I.e. to address this, you would write instead "devm_free_irq() must not
> be called within the interrupt handler because ...".
>
> > +	 * since this function waits for running interrupt handlers to finish
> > +	 * and thus it would deadlock. Instead trigger a worker that does the
> > +	 * unregistration.
> > +	 */
> > +	schedule_work(&priv->free_irq_work);
> > +}
> > +
> > +static void tpm_tis_process_unhandled_interrupt(struct tpm_chip *chip)
> > +{
> > +	const unsigned int MAX_UNHANDLED_IRQS =3D 1000;
> > +	struct tpm_tis_data *priv =3D dev_get_drvdata(&chip->dev);
>
> Reverse order and add empty line.
>
> > +	/*
> > +	 * The worker to free the TPM interrupt (free_irq_work) may already
> > +	 * be scheduled, so make sure it is not scheduled again.
> > +	 */
> > +	if (!(chip->flags & TPM_CHIP_FLAG_IRQ))
> > +		return;
> > +
> > +	if (time_after(jiffies, priv->last_unhandled_irq + HZ/10))
> > +		priv->unhandled_irqs =3D 1;
> > +	else
> > +		priv->unhandled_irqs++;
> > +
> > +	priv->last_unhandled_irq =3D jiffies;
> > +
> > +	if (priv->unhandled_irqs > MAX_UNHANDLED_IRQS)
> > +		tpm_tis_handle_irq_storm(chip);
>
> Why wouldn't we switch to polling mode even when there is a single
> unhandled IRQ?=20
>
> > +}
> > +
> >  static irqreturn_t tis_int_handler(int dummy, void *dev_id)
> >  {
> >  	struct tpm_chip *chip =3D dev_id;
> > @@ -761,10 +810,10 @@ static irqreturn_t tis_int_handler(int dummy, voi=
d *dev_id)
> > =20
> >  	rc =3D tpm_tis_read32(priv, TPM_INT_STATUS(priv->locality), &interrup=
t);
> >  	if (rc < 0)
> > -		return IRQ_NONE;
> > +		goto unhandled;
> > =20
> >  	if (interrupt =3D=3D 0)
> > -		return IRQ_NONE;
> > +		goto unhandled;
> > =20
> >  	set_bit(TPM_TIS_IRQ_TESTED, &priv->flags);
> >  	if (interrupt & TPM_INTF_DATA_AVAIL_INT)
> > @@ -780,10 +829,14 @@ static irqreturn_t tis_int_handler(int dummy, voi=
d *dev_id)
> >  	rc =3D tpm_tis_write32(priv, TPM_INT_STATUS(priv->locality), interrup=
t);
> >  	tpm_tis_relinquish_locality(chip, 0);
> >  	if (rc < 0)
> > -		return IRQ_NONE;
> > +		goto unhandled;
> > =20
> >  	tpm_tis_read32(priv, TPM_INT_STATUS(priv->locality), &interrupt);
> >  	return IRQ_HANDLED;
> > +
> > +unhandled:
> > +	tpm_tis_process_unhandled_interrupt(chip);
> > +	return IRQ_HANDLED;
> >  }
> > =20
> >  static void tpm_tis_gen_interrupt(struct tpm_chip *chip)
> > @@ -804,6 +857,15 @@ static void tpm_tis_gen_interrupt(struct tpm_chip =
*chip)
> >  		chip->flags &=3D ~TPM_CHIP_FLAG_IRQ;
> >  }
> > =20
> > +static void tpm_tis_free_irq_func(struct work_struct *work)
> > +{
> > +	struct tpm_tis_data *priv =3D container_of(work, typeof(*priv), free_=
irq_work);
> > +	struct tpm_chip *chip =3D priv->chip;
> > +
> > +	devm_free_irq(chip->dev.parent, priv->irq, chip);
> > +	priv->irq =3D 0;
> > +}
> > +
> >  /* Register the IRQ and issue a command that will cause an interrupt. =
If an
> >   * irq is seen then leave the chip setup for IRQ operation, otherwise =
reverse
> >   * everything and leave in polling mode. Returns 0 on success.
> > @@ -816,6 +878,7 @@ static int tpm_tis_probe_irq_single(struct tpm_chip=
 *chip, u32 intmask,
> >  	int rc;
> >  	u32 int_status;
> > =20
> > +	INIT_WORK(&priv->free_irq_work, tpm_tis_free_irq_func);
> > =20
> >  	rc =3D devm_request_threaded_irq(chip->dev.parent, irq, NULL,
> >  				       tis_int_handler, IRQF_ONESHOT | flags,
> > @@ -918,6 +981,7 @@ void tpm_tis_remove(struct tpm_chip *chip)
> >  		interrupt =3D 0;
> > =20
> >  	tpm_tis_write32(priv, reg, ~TPM_GLOBAL_INT_ENABLE & interrupt);
> > +	flush_work(&priv->free_irq_work);
> > =20
> >  	tpm_tis_clkrun_enable(chip, false);
> > =20
> > @@ -1021,6 +1085,7 @@ int tpm_tis_core_init(struct device *dev, struct =
tpm_tis_data *priv, int irq,
> >  	chip->timeout_b =3D msecs_to_jiffies(TIS_TIMEOUT_B_MAX);
> >  	chip->timeout_c =3D msecs_to_jiffies(TIS_TIMEOUT_C_MAX);
> >  	chip->timeout_d =3D msecs_to_jiffies(TIS_TIMEOUT_D_MAX);
> > +	priv->chip =3D chip;
> >  	priv->timeout_min =3D TPM_TIMEOUT_USECS_MIN;
> >  	priv->timeout_max =3D TPM_TIMEOUT_USECS_MAX;
> >  	priv->phy_ops =3D phy_ops;
> > diff --git a/drivers/char/tpm/tpm_tis_core.h b/drivers/char/tpm/tpm_tis=
_core.h
> > index e978f457fd4d..6fc86baa4398 100644
> > --- a/drivers/char/tpm/tpm_tis_core.h
> > +++ b/drivers/char/tpm/tpm_tis_core.h
> > @@ -91,12 +91,18 @@ enum tpm_tis_flags {
> >  };
> > =20
> >  struct tpm_tis_data {
> > +	struct tpm_chip *chip;
> >  	u16 manufacturer_id;
> >  	struct mutex locality_count_mutex;
> >  	unsigned int locality_count;
> >  	int locality;
> > +	/* Interrupts */
> >  	int irq;
> > +	struct work_struct free_irq_work;
> > +	unsigned long last_unhandled_irq;
> > +	unsigned int unhandled_irqs;
> >  	unsigned int int_mask;
> > +
> >  	unsigned long flags;
> >  	void __iomem *ilb_base_addr;
> >  	u16 clkrun_enabled;
> >
> > base-commit: 44c026a73be8038f03dbdeef028b642880cf1511
> > --=20
> > 2.40.1

I added 'irq-storm' branch where I have the latest fixes:

git://git.kernel.org/pub/scm/linux/kernel/git/jarkko/linux-tpmdd.git

All known DMI table updates are now afaik in the mainline.

BR, Jarkko