David Laight <David.Laight@xxxxxxxxxx> writes: > From: Michal Suchánek >> Sent: 09 May 2019 14:38 > ... >> > The problem is the combination of some new code called via printk(), >> > check_pointer() which calls probe_kernel_read(). That then calls >> > allow_user_access() (PPC_KUAP) and that uses mmu_has_feature() too early >> > (before we've patched features). >> >> There is early_mmu_has_feature for this case. mmu_has_feature does not >> work before patching so parts of kernel that can run before patching >> must use the early_ variant which actually runs code reading the >> feature bitmap to determine the answer. > > Does the early_ variant get patched so the it is reasonably > efficient after the 'patching' is done? No they don't get patched ever. The name is a bit misleading I guess. > Or should there be a third version which gets patched across? For a case like this it's entirely safe to just skip the code early in boot, so if it was a static_key_false everything would just work. Unfortunately the way the code is currently written we would have to change all MMU features to static_key_false and that risks breaking something else. We have a long standing TODO to rework all our feature logic and unify CPU/MMU/firmware/etc. features. Possibly as part of that we can come up with a scheme where the default value is per-feature bit. Having said all that, in this case the overhead of the test and branch is small compared to the cost of writing to the SPR which controls user access and then doing an isync, so it's all somewhat premature optimisation. cheers
