x86/mce: handle host LMCE

x86/mce: handle host LMCE

A round of mce_softirq() may handle multiple deferred MCE's.
 1/ If all of them are LMCE's, then mce_softirq() is called on one CPU
    and should not wait for others.
 2/ If at least one of them is non-local MCE, then mce_softirq()
    should sync with other CPUs. mce_softirq() should check those two
    cases and handle them accordingly.

Because mce_softirq() can be interrupted by MC# again, we should also
ensure the deferred MCE handling in mce_softirq() is immutable to the
change of the checking result.

A per-cpu list 'lmce_pending' is introduced to 'struct mc_telem_cpu_ctl'
along with the existing per-cpu list 'pending' for LMCE handling.

MC# handler mcheck_cmn_handler() ensures that
 1/ if all deferred MCE's on a CPU are LMCE's, then all of their
    telemetries will be only in 'lmce_pending' on that CPU;
 2/ if at least one of deferred MCE on a CPU is not LMCE, then all
    telemetries of deferred MCE's on that CPU will be only in
    'pending' on that CPU.

Therefore, the non-empty of 'lmce_pending' can be used to determine
whether it's the former of the beginning two cases in MCE softirq
handler mce_softirq().

mce_softirq() atomically moves deferred MCE's from either list
'lmce_pending' on the current CPU or lists 'pending' on the current or
other CPUs to list 'processing' in the current CPU, and then handles
deferred MCE's in list 'processing'.  New coming MC# before and after
the atomic move, which change the result of the check, do not change
whether MCE's in 'processing' are LMCE or not, so mce_softirq() can
still handle 'processing' according to the result of previous check.

Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>