On Tue, Jun 26, 2018 at 11:37:36AM +0900, Wolfram Sang wrote:
> On Wed, Jun 20, 2018 at 07:17:53AM +0200, Peter Rosin wrote:
> > Hi!
> > 
> > With the introduction of mux-locked I2C muxes, the concept of
> > locking only a segment of the I2C adapter tree was added. At the
> > time, I did not want to cause a lot of extra churn, so left most
> > users of i2c_lock_adapter alone and apparently didn't think enough
> > about it; they simply continued to lock the whole adapter tree.
> > However, i2c_lock_adapter is in fact wrong for almost every caller
> > (there is naturally an exception) that is itself not a driver for
> > a root adapter. What normal drivers generally want is to only
> > lock the segment of the adapter tree that their device sits on.
> > 
> > In fact, if a device sits behind a mux-locked I2C mux, and its
> > driver calls i2c_lock_adapter followed by an unlocked I2C transfer,
> > things will deadlock (since even a mux-locked I2C adapter will lock
> > its parent at some point). If the device is not sitting behind a
> > mux-locked I2C mux (i.e. either directly on the root adapter or
> > behind a (chain of) parent-locked I2C muxes) the root/segment
> > distinction is of no consequence; the root adapter is locked either
> > way.
> > 
> > Mux-locked I2C muxes are probably not that common, and putting any
> > of the affected devices behind one is probably even rarer, which
> > is why we have not seen any deadlocks. At least not that I know
> > of...
> > 
> > Since silently changing the semantics of i2c_lock_adapter might
> > be quite a surprise, especially for out-of-tree users, this series
> > instead removes the function and forces all users to explicitly
> > name I2C_LOCK_SEGMENT or I2C_LOCK_ROOT_ADAPTER in a call to
> > i2c_lock_bus, as suggested by Wolfram. Yes, users will be a teensy
> > bit more wordy, but open-coding I2C locking from random drivers
> > should be avoided, so it's perhaps a good thing if it doesn't look
> > too neat?
> > 
> > I suggest that Wolfram takes this series through the I2C tree and
> > creates an immutable branch for the other subsystems. The series
> > is based on v4.18-r1.
> 
> Applied to a seperate branch named "i2c/precise-locking-names" which I
> will merge into for-next, so it will get proper testing already. Once we
> get the missing acks from media, MFD, and IIO maintainers, I will merge
> it into for-4.19.

Ping for media related acks.

Thanks,

   Wolfram