Wolfram Sang wrote:

On Fri, Jan 08, 2016 at 04:04:48PM +0100, Peter Rosin wrote:

quoted

Hi!

[doing a v3 even if there is no "big picture" feedback yet, but
 previous versions has bugs that make them harder to test than
 needed, and testing is very much desired]

I have a pair of boards with this i2c topology:

                       GPIO ---|  ------ BAT1
                        |      v /
   I2C  -----+------B---+---- MUX
             |                   \
           EEPROM                 ------ BAT2

	(B denotes the boundary between the boards)

The problem with this is that the GPIO controller sits on the same i2c bus
that it MUXes. For pca954x devices this is worked around by using unlocked
transfers when updating the MUX. I have no such luck as the GPIO is a general
purpose IO expander and the MUX is just a random bidirectional MUX, unaware
of the fact that it is muxing an i2c bus, and extending unlocked transfers
into the GPIO subsystem is too ugly to even think about. But the general hw
approach is sane in my opinion, with the number of connections between the
two boards minimized. To put is plainly, I need support for it.

So, I observe that while it is needed to have the i2c bus locked during the
actual MUX update in order to avoid random garbage on the slave side, it
is not strictly a must to have it locked over the whole sequence of a full
select-transfer-deselect operation. The MUX itself needs to be locked, so
transfers to clients behind the mux are serialized, and the MUX needs to be
stable during all i2c traffic (otherwise individual mux slave segments
might see garbage).

This series accomplishes this by adding code to i2c-mux-gpio and
i2c-mux-pinctrl that determines if all involved devices used to update the
mux are controlled by the same root i2c adapter that is muxed. When this
is the case, the select-transfer-deselect operations should be locked
individually to avoid the deadlock. The i2c bus *is* still locked
during muxing, since the muxing happens as part of i2c transfers. This
is true even if the MUX is updated with several transfers to the GPIO (at
least as long as *all* MUX changes are using the i2s master bus). A lock
is added to the mux so that transfers through the mux are serialized.

Concerns:
- The locking is perhaps too complex?
- I worry about the priority inheritance aspect of the adapter lock. When
  the transfers behind the mux are divided into select-transfer-deselect all
  locked individually, low priority transfers get more chances to interfere
  with high priority transfers.
- When doing an i2c_transfer() in_atomic() context or with irqs_disabled(),
  there is a higher possibility that the mux is not returned to its idle
  state after a failed (-EAGAIN) transfer due to trylock.
- Is the detection of i2c-controlled gpios and pinctrls sane (i.e. the
  usage of the new i2c_root_adapter() function in 8/8)?

To summarize the series, there's some i2c-mux infrastructure cleanup work
first (I think that part stands by itself as desireable regardless), the
locking changes are in the last three patches of the series, with the real
meat in 8/8.

PS. needs a bunch of testing, I do not have access to all the involved hw

I want to let you know that I am currently thinking about this series.

Glad to hear it!

There seems to be a second occasion where it could have helped AFAICT.
http://patchwork.ozlabs.org/patch/584776/ (check my comments there from
yesterday and today)

The mpu6050 driver has to set muxc->i2c_controlled before adding
any child adapters for anything to behave differently, and it also
has to make sure that all accesses in select/deselect are normal
i2c accesses (i.e. not unlocked accesses). When doing so,
unreleated i2c traffic might interleave with the muxing. So, if
the chip is auto-deselecting on the first i2c transfer after select
it will never work properly.

First of all, really thank you that you tried to find the proper
solution and went all the way for it. It is easy to do a fire&forget
hack here, but you didn't.

Fire&forget often turns out to be just the fire. If you do it properly
there is a better chance that you really can forget it...

I hope you understand, though, that I need to make a balance between
features and complexity in my subsystem to have maintainable and stable
code.

As I wrote in the mentioned thread already: "However, I am still
undecided if that series should go upstream because it makes the mux
code another magnitude more complex. And while this seems to be the
second issue which could be fixed by that series, both issues seem to
be corner cases, so I am not sure it is worth the complexity."

And for the cleanup series using struct mux_core. It is quite an
intrusive change and, frankly, the savings look surprisingly low. I
would have expected more, but you never find out until you do it. So, I
am unsure here as well.

Yes, that part of the series went ballistic when the half-dozen mux
users outside of drivers/i2c was added to the mix (I was originally
not aware of them). The savings looked better when only the i2c-internal
muxes was considered, mainly because the external muxes are often not
real muxes and only have one child adapter. Creating the mux-core for
that one adapter then swallows the savings.

Funnily enough, I was just the other day looking at the series again and
decided to redo those 5 patches so that I first add the new mux core
and implement the old interface in terms of the new interface, then
convert all the mux users one patch at a time, then remove the glue.
That means 15 patches instead of 5, but each patch only touches one
subsystem, which should ease the transition. The end result is
equivalent, I only had to change a few names to make it possible to
have both the old and the new interfaces active at the same time.

In doing so, I realized that what might be good for the non-generic
one-child-only "muxes" is perhaps an interface that creates a mux core
and registers one child adapter with one call?

One other thing that could help the +- statistics is to maybe add more
parameters to i2c_mux_alloc, such as parent adapter and select/deselect
ops. But that is fairly cosmetic...

I am not decided and open for discussion. This is just where we are
currently. All interested parties, I am looking forward to more
thoughts.

Cheers,
Peter