]> xenbits.xensource.com Git - xen.git/commitdiff
x86/time: adjust local system time initialization
authorJan Beulich <jbeulich@suse.com>
Wed, 3 Aug 2016 12:39:31 +0000 (14:39 +0200)
committerJan Beulich <jbeulich@suse.com>
Wed, 3 Aug 2016 12:39:31 +0000 (14:39 +0200)
Using the bare return value from read_platform_stime() is not suitable
when local_time_calibration() is going to use its fast path: Divergence
of several dozen microseconds between NOW() return values on different
CPUs results when platform and local time don't stay in close sync.

Latch local and platform time on the CPU initiating AP bringup, such
that the AP can use these values to seed its stime_local_stamp with as
little of an error as possible. The boot CPU, otoh, can simply
calculate the correct initial value (other CPUs could do so too with
even greater accuracy than the approach being introduced, but that can
work only if all CPUs' TSCs start ticking at the same time, which
generally can't be assumed to be the case on multi-socket systems).

This slightly defers init_percpu_time() (moved ahead by commit
dd2658f966 ["x86/time: initialise time earlier during
start_secondary()"]) in order to reduce as much as possible the gap
between populating the stamps and consuming them.

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Tested-by: Dario Faggioli <dario.faggioli@citrix.com>
Tested-by: Joao Martins <joao.m.martins@oracle.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
xen/arch/x86/smpboot.c
xen/arch/x86/time.c
xen/include/asm-x86/time.h

index 1c02681fb6f0de081cd5bff13c63fbf946c1c223..ae99c94b39d643c56fdddb33e8ce532b680ce8b0 100644 (file)
@@ -328,12 +328,12 @@ void start_secondary(void *unused)
 
     percpu_traps_init();
 
-    init_percpu_time();
-
     cpu_init();
 
     smp_callin();
 
+    init_percpu_time();
+
     setup_secondary_APIC_clock();
 
     /*
@@ -996,6 +996,8 @@ int __cpu_up(unsigned int cpu)
     if ( (ret = do_boot_cpu(apicid, cpu)) != 0 )
         return ret;
 
+    time_latch_stamps();
+
     set_cpu_state(CPU_STATE_ONLINE);
     while ( !cpu_online(cpu) )
     {
index 9707e7688d648d39e36b985d0cbe6b3d8e2a958d..7ab3cfe18179fdd5e0c63137a0a39bda30f298a5 100644 (file)
@@ -1328,21 +1328,52 @@ static void time_calibration(void *unused)
                      &r, 1);
 }
 
+static struct {
+    s_time_t local_stime, master_stime;
+} ap_bringup_ref;
+
+void time_latch_stamps(void)
+{
+    unsigned long flags;
+    u64 tsc;
+
+    local_irq_save(flags);
+    ap_bringup_ref.master_stime = read_platform_stime();
+    tsc = rdtsc();
+    local_irq_restore(flags);
+
+    ap_bringup_ref.local_stime = get_s_time_fixed(tsc);
+}
+
 void init_percpu_time(void)
 {
     struct cpu_time *t = &this_cpu(cpu_time);
     unsigned long flags;
+    u64 tsc;
     s_time_t now;
 
     /* Initial estimate for TSC rate. */
     t->tsc_scale = per_cpu(cpu_time, 0).tsc_scale;
 
     local_irq_save(flags);
-    t->local_tsc_stamp = rdtsc();
     now = read_platform_stime();
+    tsc = rdtsc();
     local_irq_restore(flags);
 
     t->stime_master_stamp = now;
+    /*
+     * To avoid a discontinuity (TSC and platform clock can't be expected
+     * to be in perfect sync), initialization here needs to match up with
+     * local_time_calibration()'s decision whether to use its fast path.
+     */
+    if ( boot_cpu_has(X86_FEATURE_CONSTANT_TSC) )
+    {
+        if ( system_state < SYS_STATE_smp_boot )
+            now = get_s_time_fixed(tsc);
+        else
+            now += ap_bringup_ref.local_stime - ap_bringup_ref.master_stime;
+    }
+    t->local_tsc_stamp    = tsc;
     t->stime_local_stamp  = now;
 }
 
index 39d6bf3aa85bf88995ddefb673b91c2d3b62ff31..0a7cb38ec1bb5fa0acbc42e49a972dd153d0e846 100644 (file)
@@ -40,6 +40,7 @@ int time_suspend(void);
 int time_resume(void);
 
 void init_percpu_time(void);
+void time_latch_stamps(void);
 
 struct ioreq;
 int hwdom_pit_access(struct ioreq *ioreq);