I2C from interrupt context?

I2C from interrupt context?

[Krzysztof Halasa]

Hi,

I'm updating my EEPROM console logger and have encountered a problem
- the logger (as any console) can be called with hardware interrupts
disabled and/or from interrupt context. It needs to write to I^2C
(using ARM (Xscale) GPIO) and possibly SMBUS-only EEPROM chip. Is it
at all supposed to be possible?
-- 
Krzysztof Halasa

Re: I2C from interrupt context?

[David Miller]

From: Krzysztof Halasa <khc@pm.waw.pl>
Date: Fri, 07 Nov 2008 00:44:48 +0100

[ Fixed i2c list address, it's now hosted at vger ]

> I'm updating my EEPROM console logger and have encountered a problem
> - the logger (as any console) can be called with hardware interrupts
> disabled and/or from interrupt context. It needs to write to I^2C
> (using ARM (Xscale) GPIO) and possibly SMBUS-only EEPROM chip. Is it
> at all supposed to be possible?

Not really.  The I2C operations need to be able to sleep and that's
not allowed in interrupt context.

Re: I2C from interrupt context?

[Jean Delvare]

On Thu, 06 Nov 2008 15:47:54 -0800 (PST), David Miller wrote:
> From: Krzysztof Halasa <khc@pm.waw.pl>
> Date: Fri, 07 Nov 2008 00:44:48 +0100
> 
> [ Fixed i2c list address, it's now hosted at vger ]
> 
> > I'm updating my EEPROM console logger and have encountered a problem
> > - the logger (as any console) can be called with hardware interrupts
> > disabled and/or from interrupt context. It needs to write to I^2C
> > (using ARM (Xscale) GPIO) and possibly SMBUS-only EEPROM chip. Is it
> > at all supposed to be possible?
> 
> Not really.  The I2C operations need to be able to sleep and that's
> not allowed in interrupt context.

That's not totally correct. Since kernel 2.6.25, i2c_transfer()
supports being called in contexts where it cannot sleep. Of course, for
it to work, the underlying i2c adapter driver must also not sleep. As I
recall, only the i2c-pxa driver was explicitly modified to support this
at the moment (see member use_pio of struct pxa_i2c) but other drivers
could be modified in a similar way (and some i2c adapter drivers may be
naturally non-sleeping - everything based on i2c-algo-bit is likely to
fall into this category.)

The situation is far from perfect though. For one thing, I seem to
recall that Andrew Morton didn't like the approach taken in
i2c_transfer(). For another, i2c_smbus_xfer() was not yet modified so
at this point only I2C-level transactions can be non-sleeping,
SMBus-level transactions can't. But all this could be fixed by anyone
who cares about these specific issues.

-- 
Jean Delvare

Re: I2C from interrupt context?

[Krzysztof Halasa]

Jean Delvare <khali@linux-fr.org> writes:

> The situation is far from perfect though. For one thing, I seem to
> recall that Andrew Morton didn't like the approach taken in
> i2c_transfer(). For another, i2c_smbus_xfer() was not yet modified so
> at this point only I2C-level transactions can be non-sleeping,
> SMBus-level transactions can't. But all this could be fixed by anyone
> who cares about these specific issues.

Thanks, I'll look at it.
-- 
Krzysztof Halasa

Re: I2C from interrupt context?

[Andrew Morton]

On Fri, 07 Nov 2008 17:05:35 +0100 Krzysztof Halasa <khc@pm.waw.pl> wrote:

> Jean Delvare <khali@linux-fr.org> writes:
> 
> > The situation is far from perfect though. For one thing, I seem to
> > recall that Andrew Morton didn't like the approach taken in
> > i2c_transfer(). For another, i2c_smbus_xfer() was not yet modified so
> > at this point only I2C-level transactions can be non-sleeping,
> > SMBus-level transactions can't. But all this could be fixed by anyone
> > who cares about these specific issues.
> 
> Thanks, I'll look at it.

The problem (well: bug) is that in_atomic() returns false inside a
spinlock when CONFIG_PREEMPT=n.  The code as it stands can sleep
inside a spinlock, which is deadlockable if a scheduled-to task
tries to take the same spinlock.

There is no means like this by which a piece of code can determine
whether it can call schedule().  The pattern which we use in many many
places (most especially GFP_KERNEL/GFP_ATOMIC) is to pass a flag down
to callees telling them in some manner which context they were called from.