It appears there is a regression for 32-bit kernels due to SME changes.
I bisected my particular problem (Xen PV guest) to 21729f81ce8ae76a6995681d40e16f7ce8075db4 but I also saw pmd_clear_bad() errors on baremetal. This seems to be caused by sme_me_mask being an unsigned long as opposed to phys_addr_t (the actual problem is that __PHYSICAL_MASK is truncated). When I declare it as u64 and drop unsigned long cast in __sme_set()/__sme_clr() the problem goes way. (This presumably won't work for non-PAE which I haven't tried).
-boris
