x86/tsc: Avoid synchronizing TSCs with multiple CPUs in parallel

The TSC sync algorithm is only designed to do a 1:1 sync between the
source and target CPUs.

In order to enable parallel CPU bringup, serialize it by using an
atomic_t containing the number of the target CPU whose turn it is.

In future we should look at inventing a 1:many TSC synchronization
algorithm, perhaps falling back to 1:1 if a warp is observed but
doing them all in parallel for the common case where no adjustment
is needed. Or just avoiding the sync completely for cases like kexec
where we trust that they were in sync already.

This is perfectly sufficient for the short term though, until we get
those further optimisations.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>

diff --git a/arch/x86/kernel/tsc_sync.c b/arch/x86/kernel/tsc_sync.c
index 9452dc9664b51fddcfaeb6274935c91885814fda..ee9278edc96ff1d84b6fcc68def7217c88f91b75 100644 (file)
--- a/arch/x86/kernel/tsc_sync.c
+++ b/arch/x86/kernel/tsc_sync.c
@@ -243,6 +243,7 @@ bool tsc_store_and_check_tsc_adjust(bool bootcpu)
  * Entry/exit counters that make sure that both CPUs
  * run the measurement code at once:
  */
+static atomic_t tsc_sync_cpu = ATOMIC_INIT(-1);
 static atomic_t start_count;
 static atomic_t stop_count;
 static atomic_t skip_test;
@@ -367,6 +368,8 @@ void check_tsc_sync_source(int cpu)
                atomic_set(&test_runs, 1);
        else
                atomic_set(&test_runs, 3);
+
+       atomic_set(&tsc_sync_cpu, cpu);
 retry:
        /*
         * Wait for the target to start or to skip the test:
@@ -448,6 +451,10 @@ void check_tsc_sync_target(void)
        if (unsynchronized_tsc())
                return;
 
+       /* Wait for this CPU's turn */
+       while (atomic_read(&tsc_sync_cpu) != cpu)
+               cpu_relax();
+
        /*
         * Store, verify and sanitize the TSC adjust register. If
         * successful skip the test.