Hi Adrian: * Adrian Cox <adrian at humboldt.co.uk> [2004-09-22 19:32:31 +0100]: > On Wed, 2004-09-22 at 14:38, Jean Delvare wrote: > > > Aha, this is an interesting point (which was missing from your previous > > explanation). The base of your proposal would be to have several small i2c > > "trees" (where a tree is a list of adapters and a list of clients) instead of > > a larger, unique one. This would indeed solve a number of problems, and I > > admit that it is somehow equivalent to Michael's classes in that it > > efficiently prevents the hardware monitoring clients from probing the video > > stuff. The rest is just details internal to each "tree". As I understand it, > > each video device would be a tree on itself, while the whole hardware > > monitoring stuff would constitute one (bigger) tree. Correct? > > I've been rereading the code, and it could be even simpler. How about > this: > > 1) The card driver defines an i2c_adapter structure, but never calls > i2c_add_adapter(). The only extra thing it needs to do is to initialise > the semaphores in the structure. > 2) The frontend calls i2c_transfer() directly. > 3) The i2c core never gets involved, and there is never any i2c_client > structure. Yes, almost... Why force your card driver to re-implement i2c_smbus_read_byte() and all its relatives? Go ahead and define the i2c_client structure(s) as well, but don't i2c_attach_client(). Sensors drivers do their probing before attaching the client, so I know that works. > This gives us the required reuse of the I2C algo-bit code, without any > of the list walking or device probing being required. Ditto, *plus* you can still reuse all the i2c_core helper routines that require a client structure. Regards, -- Mark M. Hoffman mhoffman at lightlink.com
