arch/tile: fix __ndelay etc to work better

The current implementations of __ndelay and __udelay call a hypervisor
service to delay, but the hypervisor service isn't actually implemented
very well, and the consensus is that Linux should handle figuring this
out natively and not use a hypervisor service.

By converting nanoseconds to cycles, and then spinning until the
cycle counter reaches the desired cycle, we get several benefits:
first, we are sensitive to the actual clock speed; second, we use
less power by issuing a slow SPR read once every six cycles while
we delay; and third, we properly handle the case of an interrupt by
exiting at the target time rather than after some number of cycles.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>
diff --git a/arch/tile/include/asm/timex.h b/arch/tile/include/asm/timex.h
index 3baf5fc..29921f0 100644
--- a/arch/tile/include/asm/timex.h
+++ b/arch/tile/include/asm/timex.h
@@ -38,6 +38,9 @@
 
 cycles_t get_clock_rate(void);
 
+/* Convert nanoseconds to core clock cycles. */
+cycles_t ns2cycles(unsigned long nsecs);
+
 /* Called at cpu initialization to set some low-level constants. */
 void setup_clock(void);
 
diff --git a/arch/tile/include/hv/hypervisor.h b/arch/tile/include/hv/hypervisor.h
index f672544..103986b 100644
--- a/arch/tile/include/hv/hypervisor.h
+++ b/arch/tile/include/hv/hypervisor.h
@@ -964,6 +964,11 @@
 
 /** Waits for at least the specified number of nanoseconds then returns.
  *
+ * NOTE: this deprecated function currently assumes a 750 MHz clock,
+ * and is thus not generally suitable for use.  New code should call
+ * hv_sysconf(HV_SYSCONF_CPU_SPEED), compute a cycle count to wait for,
+ * and delay by looping while checking the cycle counter SPR.
+ *
  * @param nanosecs The number of nanoseconds to sleep.
  */
 void hv_nanosleep(int nanosecs);
diff --git a/arch/tile/kernel/entry.S b/arch/tile/kernel/entry.S
index fd8dc42..c3aa067 100644
--- a/arch/tile/kernel/entry.S
+++ b/arch/tile/kernel/entry.S
@@ -38,12 +38,6 @@
 	jrp lr
 	STD_ENDPROC(kernel_execve)
 
-/* Delay a fixed number of cycles. */
-STD_ENTRY(__delay)
-	{ addi r0, r0, -1; bnzt r0, . }
-	jrp lr
-	STD_ENDPROC(__delay)
-
 /*
  * We don't run this function directly, but instead copy it to a page
  * we map into every user process.  See vdso_setup().
diff --git a/arch/tile/kernel/time.c b/arch/tile/kernel/time.c
index f2e156e..49a605b 100644
--- a/arch/tile/kernel/time.c
+++ b/arch/tile/kernel/time.c
@@ -224,3 +224,13 @@
 {
 	return -EINVAL;
 }
+
+/*
+ * Use the tile timer to convert nsecs to core clock cycles, relying
+ * on it having the same frequency as SPR_CYCLE.
+ */
+cycles_t ns2cycles(unsigned long nsecs)
+{
+	struct clock_event_device *dev = &__get_cpu_var(tile_timer);
+	return ((u64)nsecs * dev->mult) >> dev->shift;
+}
diff --git a/arch/tile/lib/delay.c b/arch/tile/lib/delay.c
index 5801b03..cdacdd1 100644
--- a/arch/tile/lib/delay.c
+++ b/arch/tile/lib/delay.c
@@ -15,20 +15,31 @@
 #include <linux/module.h>
 #include <linux/delay.h>
 #include <linux/thread_info.h>
-#include <asm/fixmap.h>
-#include <hv/hypervisor.h>
+#include <asm/timex.h>
 
 void __udelay(unsigned long usecs)
 {
-	hv_nanosleep(usecs * 1000);
+	if (usecs > ULONG_MAX / 1000) {
+		WARN_ON_ONCE(usecs > ULONG_MAX / 1000);
+		usecs = ULONG_MAX / 1000;
+	}
+	__ndelay(usecs * 1000);
 }
 EXPORT_SYMBOL(__udelay);
 
 void __ndelay(unsigned long nsecs)
 {
-	hv_nanosleep(nsecs);
+	cycles_t target = get_cycles();
+	target += ns2cycles(nsecs);
+	while (get_cycles() < target)
+		cpu_relax();
 }
 EXPORT_SYMBOL(__ndelay);
 
-/* FIXME: should be declared in a header somewhere. */
+void __delay(unsigned long cycles)
+{
+	cycles_t target = get_cycles() + cycles;
+	while (get_cycles() < target)
+		cpu_relax();
+}
 EXPORT_SYMBOL(__delay);