lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAPY8ntCHde9hXdvsGCNrmHxEmBME_=yYWquq0HVNoo6EDJL6hQ@mail.gmail.com>
Date:   Tue, 31 Oct 2023 18:04:32 +0000
From:   Dave Stevenson <dave.stevenson@...pberrypi.com>
To:     Mike Isely at pobox <isely@...ox.com>
Cc:     Stefan Wahren <wahrenst@....net>, mike.isely@...altdigital.com,
        Andi Shyti <andi.shyti@...nel.org>,
        Florian Fainelli <florian.fainelli@...adcom.com>,
        Phil Elwell <phil@...pberrypi.com>,
        Broadcom internal kernel review list 
        <bcm-kernel-feedback-list@...adcom.com>,
        Ray Jui <rjui@...adcom.com>,
        Scott Branden <sbranden@...adcom.com>,
        linux-rpi-kernel@...ts.infradead.org,
        linux-arm-kernel@...ts.infradead.org,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 2/2] [i2c-bcm2835] ALWAYS enable INTD

Hi Mike

On Tue, 31 Oct 2023 at 16:44, Mike Isely <isely@...ly.net> wrote:
>
> On Tue, 31 Oct 2023, Dave Stevenson wrote:
>
> > (Thanks Stefan for forwarding - linux-rpi-kernel seems to be a little
> > too aggressive on spam filtering, and I'm not on the other lists
> > cc'ed).
> >
> > Hi Mike
> >
> > On Tue, 31 Oct 2023 at 12:37, Stefan Wahren <wahrenst@....net> wrote:
> > >
> > > [Forware to Dave and Phil]
> > >
> > > Am 30.10.23 um 17:21 schrieb mike.isely@...altdigital.com:
> > > > From: Mike Isely <mike.isely@...altdigital.com>
> > > >
> > > > There is a race in the bcm2835 i2c hardware: When one starts a write
> > > > transaction, two things apparently take place at the same time: (1) an
> > > > interrupt is posted to cause the FIFO to be filled with TX data,
> > > > and (2) an I2C transaction is started on the wire with the slave
> > > > select byte.  The race happens if there's no slave, as this causes a
> > > > slave selection timeout, raising the ERR flag in the hardware and
> > > > setting DONE.  The setting of that DONE flag races against TXW.  If
> > > > TXW gets set first, then an interrupt is raised if INTT was set.  If
> > > > ERR gets set first, then an interrupt is raised if INTD was set.  It's
> > > > one or the other, not both - probably because DONE being set disables
> > > > the hardware INTT interrupt path.
> >
> > I'm not following the full sequence of events required here.
> > If you only had a slave selection message, then num_msgs = 1 and INTD
> > will be enabled immediately anyway.
>
> For an I2C read, there typically needs to first be a write, thus always
> two messages.  This is of course the case for any I2C EEPROM read.  If

Terminology.
Your mention of "slave select byte" is referencing writing an address
to the EEPROM, not sending the I2C address which is also slave
selection and one byte long.
You get a NAK on sending that EEPROM address selection byte due to the
device not being present, not a specific error from the controller
that is a "slave select timeout".

This is where my confusion comes from. Please use the I2C terminology.

> the target is not present, then the slave select error happens on the
> first message and since INTD only gets set on the second message in this
> case, then you have the setup for the race (see my other reply).  If the
> slave select "wins" instead of the TX interrupt, it sets DONE along with
> ERR and I'm guessing that in doing so the TX interrupt is masked,
> blocking it.  Without INTD also enabled then the ISR never fires - and
> we hang.
>
> What I have observed empirically is that this "almost" never happens.
> Almost.  For an I2C write in that first message there should always be a
> TX interrupt to start filling the FIFO and the ISR in that case is
> noticing ERR set along the way.  But if the ERR is set before the TX
> interrupt, then we're in trouble.  It's rare but it does happen.
>
> Note also that this is only and issue when accessing a missing target.
> For normal access this can't ever happen.  So the scenario itself is
> also rare - normally you would only cause this by running a bus scan
> (i2cdetect) at the CLI.  In our case the I2C target in question is a
> device in an SFP cage that may or may not be populated, and when not
> populated apparently the Linux kernel sfp module polls it, making this
> rainy day scenario much more likely.

OK, at least we now know your scenario and can try to replicate.

I've never seen ERR set before the end of the I2C transaction. Are you
sure this is definitely triggered by your main transaction, and isn't
left around from a previous one as you've alluded to?
Potentially it is this patch that allows ERR to be signalled earlier
as the DONE interrupt is unmasked.

> > I did investigate some I2C issues back in May due to observed issues
> > between one of the camera modules and the DSI screen touch controller.
> > If memory serves correctly, the biggest issue I found was that
> > aborting the transaction when active just left SDA & SCL in whatever
> > was the current state, including midway through a byte and no stop
> > condition. I didn't find a valid way to do a controlled stop, and
> > therefore ended up with a patch that will always complete the
> > transaction before looking at the status flags [1]. (Yes, I really
> > should upstream those patches).
>
> The key here is that we're actually not aborting the transaction, as far
> as the silicon is concerned.  What's happening here is that the silicon
> completed its operation, setting the DONE bit as one might expect, but
> we never get the ISR due to the race in the hardware.  Thus, a timeout
> happens.  Realize this is not a timeout on the I2C bus but a timeout
> between the driver and the I2C controller.  The I2C bus is idle as it
> should be.
>
> Given that the hardware can detect clock stretch problems then there
> really should be no circumstance where the controller silicon can leave
> the bus stuck.

I can't find my debug notes from May, but it was aborting part way
through a multi-byte read from the touchscreen without a STOP
condition, leaving the touch controller still thinking it was still
clocking out the current byte, so it messed up the next transaction.
Hitting BCM2835_I2C_C with BCM2835_I2C_C_CLEAR whilst BCM2835_I2C_S_TA
is set will abort immediately with no STOP :-(

>
> >
> > For a linked thread[2] I think I found that the ERR flag wasn't
> > signalled until the end of the complete transaction.
>
> That may be the case but we're talking about 2 transactions, a write and
> a read.  If INTD is only set on the second transaction then we risk a
> missed interrupt.
>
> The hardware doesn't count two messages and two transactions.  The
> driver has to set each one up as a separate transaction.  So your point
> is still correct, but from the POV of the driver it's at the end of the
> *message* not the *transaction*.
>
>
> >
> >
> > Hang on, if you're always enabling BCM2835_I2C_C_INTD, then if we have
> > a write of N bytes and read of M bytes, don't we get a DONE after the
> > write, meaning that the ISR completes then due to the clause at [3]
> > and we never do the read? Something feels wrong here.
>
> The write and the read are separate transactions as far as the hardware
> is concerned.

Yes and no. You keep the hardware busy so that you get a repeated
start, rather than a stop and new start for two totally independent
I2C operations.

>  DONE means the end of the message not the end of the
> entire transaction.  I think there's a terminology disconnect here:

DONE is meant to signal the end of a transfer, but does this restart
stop it sending the DONE? I really don't know, and don't immediately
have the time to investigate.

> Hardware: transaction == a single message
> Driver: transaction == all messages combined.
>
> Hmm, looking further at the ISR I see it's going to short-circuit the
> entire transaction (driver P.O.V.) if DONE is set.  All I can say about
> that is that with this change, everything *IS* working properly here.
> Full EEPROM access is functioning fine and that requires 2
> messages / transaction in the driver.  If DONE were short-circuiting
> this, then that would break and it isn't.

This is my concern, and it'll need a lot of testing to confirm
behaviour. (I doubt we can get Broadcom folks to dig into the RTL to
confirm what the hardware actually does as the controller is obsolete
and the designers have all left many years back).

What happens with multiple writes in one master_xfer call? Or a single
write or read?
EEPROM access is one small subset of I2C access patterns, so it needs
validating in the other potential ones.


When Phil investigated part of [1] to abort on ERR at the end of an
i2c_msg (not complete master_xfer call), he had also observed that
enabling INTD did work. However he caveated it with needing to prefill
the FIFO and handle DONE appropriate in the ISR as it no longer means
all complete. You've done the first of those three, so I'm wanting his
view on this.

  Dave

[1] https://github.com/raspberrypi/linux/issues/5429#issuecomment-1508364342

> I can probably instrument things further to determine what is really
> going on with DONE in this case.  There has to be a better explanation.
>
>   -Mike
>
>
> >
> >   Dave
> >
> > [1] https://github.com/raspberrypi/linux/pull/5479/commits
> > [2] https://forums.raspberrypi.com/viewtopic.php?p=2098691#p2098691
> > [3] https://github.com/torvalds/linux/blob/master/drivers/i2c/busses/i2c-bcm2835.c#L293-L306
> >
> > > >
> > > > MOST of the time, TXW gets set first, the ISR runs, sees ERR is set
> > > > and cleanly fails the transaction.  However some of the time DONE gets
> > > > set first - but since the driver doesn't enable INTD until it's on the
> > > > last message - there's no interrupt at all.  Thus the ISR never fires
> > > > and the driver detects a timeout instead.  At best, the "wrong" error
> > > > code is delivered to the owner of the transaction.  At worst, if the
> > > > timeout doesn't propertly clean up the hardware (see prior commit
> > > > fixing that), the next - likely unrelated - transaction will get
> > > > fouled, leading to bizarre behavior in logic otherwise unrelated to
> > > > the source of the original error.
> > > >
> > > > The fix here is to set INTD on for all messages not just the last one.
> > > > In that way, unexpected failures which might set DONE earlier than
> > > > expected will always trigger an interrupt and be handled correctly.
> > > >
> > > > The datasheet for this hardware doesn't describe any scenario where
> > > > the hardware can realistically hang - even a stretched clock will be
> > > > noticed if it takes too long.  So in theory a timeout should really
> > > > NEVER happen, and with this fix I was completely unable to trigger any
> > > > further timeouts at all.
> > > >
> > > > Signed-off-by: Mike Isely <isely@...ox.com>
> > > > ---
> > > >   drivers/i2c/busses/i2c-bcm2835.c | 6 +-----
> > > >   1 file changed, 1 insertion(+), 5 deletions(-)
> > > >
> > > > diff --git a/drivers/i2c/busses/i2c-bcm2835.c b/drivers/i2c/busses/i2c-bcm2835.c
> > > > index 96de875394e1..70005c037ff9 100644
> > > > --- a/drivers/i2c/busses/i2c-bcm2835.c
> > > > +++ b/drivers/i2c/busses/i2c-bcm2835.c
> > > > @@ -235,26 +235,22 @@ static void bcm2835_drain_rxfifo(struct bcm2835_i2c_dev *i2c_dev)
> > > >
> > > >   static void bcm2835_i2c_start_transfer(struct bcm2835_i2c_dev *i2c_dev)
> > > >   {
> > > > -     u32 c = BCM2835_I2C_C_ST | BCM2835_I2C_C_I2CEN;
> > > > +     u32 c = BCM2835_I2C_C_ST | BCM2835_I2C_C_I2CEN | BCM2835_I2C_C_INTD;
> > > >       struct i2c_msg *msg = i2c_dev->curr_msg;
> > > > -     bool last_msg = (i2c_dev->num_msgs == 1);
> > > >
> > > >       if (!i2c_dev->num_msgs)
> > > >               return;
> > > >
> > > >       i2c_dev->num_msgs--;
> > > >       i2c_dev->msg_buf = msg->buf;
> > > >       i2c_dev->msg_buf_remaining = msg->len;
> > > >
> > > >       if (msg->flags & I2C_M_RD)
> > > >               c |= BCM2835_I2C_C_READ | BCM2835_I2C_C_INTR;
> > > >       else
> > > >               c |= BCM2835_I2C_C_INTT;
> > > >
> > > > -     if (last_msg)
> > > > -             c |= BCM2835_I2C_C_INTD;
> > > > -
> > > >       bcm2835_i2c_writel(i2c_dev, BCM2835_I2C_A, msg->addr);
> > > >       bcm2835_i2c_writel(i2c_dev, BCM2835_I2C_DLEN, msg->len);
> > > >       bcm2835_i2c_writel(i2c_dev, BCM2835_I2C_C, c);
> > > >   }
> > >
> >

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ