Re: [PATCH 0/2] hw/i2c: Adds pca954x i2c mux switch device

[Corey Minyard <minyard@acm.org>]

On Tue, Apr 06, 2021 at 08:55:14AM -0700, Patrick Venture wrote:
> On Tue, Apr 6, 2021 at 8:41 AM Patrick Venture <venture@google.com> wrote:
> >
> > On Mon, Apr 5, 2021 at 12:58 PM Corey Minyard <cminyard@mvista.com> wrote:
> > >
> > > On Sat, Apr 03, 2021 at 03:28:08PM -0700, Patrick Venture wrote:
> > > > The i2c mux device pca954x implements two devices:
> > > >  - the pca9546 and pca9548.
> > > >
> > > > Patrick Venture (2):
> > > >   hw/i2c/core: add reachable state boolean
> > > >   hw/i2c: add pca954x i2c-mux switch
> > >
> > > Looking this over, the code looks good, but I have a few general
> > > questions:
> > >
> > > * Can you register the same slave address on different channels?  That's
> > >   something you could do with real hardware and might be required at
> > >   some time.  It looks like to me that you can't with this patch set,
> > >   but maybe I'm missing something.
> >
> > If I understand the hardware's implementation properly you can have
> > collisions, and this allows for collisions.  I'm not sure what you
> > mean by having both accessible.  For instance, on hardware you can
> > have a switch with N channels, and on two of the channels there is an
> > eeprom at 50.  But you're unable to talk to both eeproms at the same
> > time, because the addresses collide -- so how would the hardware know
> > which you're talking to?  My understanding of the behavior in this
> > collision case is that it just talks to the first one that responds
> > and can lead to unexpected things.
> >
> > There is a board, the quanta-q71l where we had to set the
> > idle-disconnect because there were two muxes on the same bus, with
> > conflicting addresses, and so we had to use idle disconnect explicitly
> > to make the software happy talking to the hardware -- not ideal as
> > having two devices behind different channels, but ultimately it's the
> > same idea because the devices are conflicting.
> >
> > >
> > > * Can you add devices to the secondary I2C busses on the mux using the
> > >   standard QEMU device model, or is the function call required?
> >
> > I added the function call because I didn't see a clean way to bridge
> > the issue as well as, the quasi-arbitrary bus numbering used by the
> > kernel isn't how the hardware truly behaves, and my goal was to
> > implement closer to the hardware.  I thought about adding an I2cBus to
> > the device and then you'd be able to access it, but wasn't sure of a
> > nice clean way to plumb that through -- I considered adding/removing
> > devices from the parent i2c bus instead of the boolean reachable, but
> > that seemed way less clean - although do-able.
> >
> > >
> > > I ask because I did a pca9540 and pca9541 device, but I've never
> > > submitted it because I didn't think it would ever be needed.  It takes a
> > > different tack on the problem; it creates the secondary busses as
> > > standard QEMU I2C busses and bridges them.  You can see it at
> > >
> > >    github.com:cminyard/qemu.git master-i2c-rebase
> > >
> >
> > I'll have to take a look at your approach, but the idea that it
> > wouldn't be needed sounds bizarre to me as nearly all BMC-based qemu
> > boards leverage i2c muxes to handle their PCIe slot i2c routing.
> >
> > > If you design can do the things I ask, then it's better.  If not, then
> > > I'm not sure.
> 
> Corey,
> 
> looking at your design, I should be able to do something similar with
> a small tweak.
> 
> I think my design follows the hardware where there can be conflicts,
> etc, but what I didn't know how to do was add the faux I2cBuses in a
> useful way -- but if I add the I2cBuses to the device, and then on
> add/remove it registers the device on the parent bus -- i can still
> use the reachable boolean to control whether it's present.  The faux
> I2cBuses would be a simplification for adding/removing i2c devices --
> and would act as the device list in my object.  So then setting the
> channels would change to walking the devices held by the bus that
> corresponds with the bit -- but _still_ using the reachable boolean.
> 
> If you'd like, I can update my patchset to use an i2cbus for the
> purpose above, then it would satisfy the requirement of leveraging the
> normal device process and no longer require the special function call.

That sounds reasonable.  Your implementation is quite a bit simpler than
mine, which is a bonus.

-corey

> 
> Patrick
> 
> > >
> > > -corey
> > >
> > > >
> > > >  MAINTAINERS                      |   6 +
> > > >  hw/i2c/Kconfig                   |   4 +
> > > >  hw/i2c/core.c                    |   6 +
> > > >  hw/i2c/i2c_mux_pca954x.c         | 182 +++++++++++++++++++++++++++++++
> > > >  hw/i2c/meson.build               |   1 +
> > > >  hw/i2c/trace-events              |   5 +
> > > >  include/hw/i2c/i2c.h             |   3 +
> > > >  include/hw/i2c/i2c_mux_pca954x.h |  60 ++++++++++
> > > >  8 files changed, 267 insertions(+)
> > > >  create mode 100644 hw/i2c/i2c_mux_pca954x.c
> > > >  create mode 100644 include/hw/i2c/i2c_mux_pca954x.h
> > > >
> > > > --
> > > > 2.31.0.208.g409f899ff0-goog
> > > >
>