When you change a page, you basically can access only the registers in
this page so if there are two functions requesting different pages at
the same time or registers of different pages, it won't work well
indeed.

quoted

phy_read_page() and phy_write_page() will do the needed locking if
this is an issue.

That's awesome! Didn't know it existed. Thanks a ton!

Well, that means I should migrate the whole driver to use
phy_read/write_paged instead of the phy_read/write that is currently in
use.

That's impacting performance though as per phy_read/write_paged we read
the current page, set the desired page, read/write the register, set the
old page back. That's 4 times more operations.

You can use the lower level locking primatives. See m88e1318_set_wol()
for example.

Couldn't we use the
phy_device mutex instead (as it's currently done in the whole driver)?
Or is it worse/comparable in performance to the suggested solution?

Russell King found a race condition where this breaks. You cannot hold
the phy_device mutex everywhere.

    Andrew