[PATCH] ppc: improve timebase sync for SMP

[Paul Mackerras <paulus@samba.org>]

Currently the procedure in the ppc32 kernel that synchronizes the
timebase registers across an SMP powermac system does so by setting
both timebases to zero.  That is OK at boot but causes problems if
done later.  So that we can do hotplug CPU on these machines, this
patch changes the code so it reads the timebase from one CPU and
transfers the value to the other CPU.  (Hotplug CPU is needed for
sleep (aka suspend to RAM) to work.)

Signed-off-by: Paul Mackerras <paulus@samba.org>

diff -urN linux-2.5/arch/ppc/platforms/pmac_smp.c pmac-2.5/arch/ppc/platforms/pmac_smp.c
--- linux-2.5/arch/ppc/platforms/pmac_smp.c	2005-03-15 10:18:23.000000000 +1100
+++ pmac-2.5/arch/ppc/platforms/pmac_smp.c	2005-03-15 11:59:02.000000000 +1100
@@ -116,6 +116,8 @@
 
 /* Sync flag for HW tb sync */
 static volatile int sec_tb_reset = 0;
+static unsigned int pri_tb_hi, pri_tb_lo;
+static unsigned int pri_tb_stamp;
 
 static void __init core99_init_caches(int cpu)
 {
@@ -453,7 +455,7 @@
 #endif
 	struct device_node *cpus, *firstcpu;
 	int i, ncpus = 0, boot_cpu = -1;
-	u32 *tbprop;
+	u32 *tbprop = NULL;
 
 	if (ppc_md.progress) ppc_md.progress("smp_core99_probe", 0x345);
 	cpus = firstcpu = find_type_devices("cpu");
@@ -576,46 +578,74 @@
 	}
 }
 
-void __init smp_core99_take_timebase(void)
+/* not __init, called in sleep/wakeup code */
+void smp_core99_take_timebase(void)
 {
-	/* Secondary processor "takes" the timebase by freezing
-	 * it, resetting its local TB and telling CPU 0 to go on
-	 */
-	pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 4);
-	pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
+	unsigned long flags;
+
+	/* tell the primary we're here */
+	sec_tb_reset = 1;
 	mb();
 
-	set_dec(tb_ticks_per_jiffy);
-	set_tb(0, 0);
-	last_jiffy_stamp(smp_processor_id()) = 0;
+	/* wait for the primary to set pri_tb_hi/lo */
+	while (sec_tb_reset < 2)
+		mb();
 
+	/* set our stuff the same as the primary */
+	local_irq_save(flags);
+	set_dec(1);
+	set_tb(pri_tb_hi, pri_tb_lo);
+	last_jiffy_stamp(smp_processor_id()) = pri_tb_stamp;
+	mb();
+
+	/* tell the primary we're done */
+       	sec_tb_reset = 0;
 	mb();
-       	sec_tb_reset = 1;
+	local_irq_restore(flags);
 }
 
-void __init smp_core99_give_timebase(void)
+/* not __init, called in sleep/wakeup code */
+void smp_core99_give_timebase(void)
 {
+	unsigned long flags;
 	unsigned int t;
 
-	/* Primary processor waits for secondary to have frozen
-	 * the timebase, resets local TB, and kick timebase again
-	 */
-	/* wait for the secondary to have reset its TB before proceeding */
-	for (t = 1000; t > 0 && !sec_tb_reset; --t)
-		udelay(1000);
-	if (t == 0)
+	/* wait for the secondary to be in take_timebase */
+	for (t = 100000; t > 0 && !sec_tb_reset; --t)
+		udelay(10);
+	if (!sec_tb_reset) {
 		printk(KERN_WARNING "Timeout waiting sync on second CPU\n");
+		return;
+	}
 
-       	set_dec(tb_ticks_per_jiffy);
-	set_tb(0, 0);
-	last_jiffy_stamp(smp_processor_id()) = 0;
+	/* freeze the timebase and read it */
+	/* disable interrupts so the timebase is disabled for the
+	   shortest possible time */
+	local_irq_save(flags);
+	pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 4);
+	pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
+	mb();
+	pri_tb_hi = get_tbu();
+	pri_tb_lo = get_tbl();
+	pri_tb_stamp = last_jiffy_stamp(smp_processor_id());
 	mb();
 
+	/* tell the secondary we're ready */
+	sec_tb_reset = 2;
+	mb();
+
+	/* wait for the secondary to have taken it */
+	for (t = 100000; t > 0 && sec_tb_reset; --t)
+		udelay(10);
+	if (sec_tb_reset)
+		printk(KERN_WARNING "Timeout waiting sync(2) on second CPU\n");
+	else
+		smp_tb_synchronized = 1;
+
 	/* Now, restart the timebase by leaving the GPIO to an open collector */
        	pmac_call_feature(PMAC_FTR_WRITE_GPIO, NULL, core99_tb_gpio, 0);
         pmac_call_feature(PMAC_FTR_READ_GPIO, NULL, core99_tb_gpio, 0);
-
-	smp_tb_synchronized = 1;
+	local_irq_restore(flags);
 }
 
 
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/