A set of x86 fixes:

  - Fix 32-bit PTI for real. pti_clone_entry_text() is called twice, once
    before initcalls so that initcalls can use the user-mode helper and
    then again after text is set read only. Setting read only on 32-bit
    might break up the PMD mapping, which makes the second invocation of
    pti_clone_entry_text() find the mappings out of sync and failing.

    Allow the second call to split the existing PMDs in the user mapping
    and synchronize with the kernel mapping.

  - Don't make acpi_mp_wake_mailbox not read only after init as the mail
    box must be writable in the case that CPU hotplug operations happen
    after boot. Otherwise the attempt to start a CPU crashes with a write
    to read only memory.

  - Add a missing sanity check in mtrr_save_state() to ensure that the
    fixed MTRR MSRs are supported.

    Otherwise mtrr_save_state() ends up in a #GP, which is fixed up, but
    the WARN_ON() can bring systems down when panic on warn is set.